The Federal Trade Commission has ordered Delta Air Lines, United Airlines and JetBlue, alongside Uber, Lyft, DoorDash, Instacart, Target and Walmart, to hand over detailed data on how they set prices and, for the gig platforms, how they pay workers. The orders were issued under Section 6(b) of the FTC Act. If a company fails to turn over the required information, the Commission has the power to compel compliance in federal court.

Delta disclosed in 2024 that it was testing AI-assisted fare technology built with a firm called Fetcherr. In responses to the new inquiry, Delta and JetBlue denied using personal data to set individualised fares, with Delta saying it had never used, tested or planned to use a product that targets customers with personalised offers based on personal data.

“Consumers deserve to know whether companies use their web surfing habits or other personal data”

That fragment, from FTC Chairman Andrew N. Ferguson, frames the inquiry as a correction to a 2024 study that questioned pricing consultancies rather than the companies actually setting prices. That study examined only middlemen consultancies rather than the businesses that actually set prices for services and products offered to consumers, and the agency learned nothing about actual pricing decisions from it. Going straight to the companies is the more useful approach for finding out what the algorithms do, rather than what the industry selling them claims.

Nobody has shown that any of the nine companies charges two customers different prices for the same seat or cart based on who they are, and the airlines have been quick to deny it. What the inquiry tests is whether the tools now available to do that quietly, at scale, are being used that way, and whether anyone outside the companies would otherwise know.

The New York Times filed a lawsuit against Perplexity on the fifth of December, saying the AI search company copied its journalism to deliver it to customers “without permission or compensation”. The Chicago Tribune had filed its own suit a day earlier, on the fourth of December, alleging Perplexity unlawfully copied millions of its stories, videos and images to feed its consumer answer engine, enterprise chatbot, its application programming interfaces and its Comet browser.

Two newsrooms reaching the same legal conclusion within a day of each other points to a shared read of where settlement talks with Perplexity have stalled, after many other publishers chose licensing deals instead.

Two lawsuits filed a day apart read less like coincidence and more like a shared verdict on where talks with Perplexity went

Perplexity’s business depends on summarising and answering from material it does not own, a premise that has already drawn a separate Amazon suit over its shopping agent and years of publisher warnings that answer engines strip away the traffic that once funded the reporting itself. The Times and Tribune complaints allege systematic copying across each paper’s entire archive, delivered through a named consumer browser as well as backend software that readers never see, rather than a dispute over a single contested output.

Nothing here has been decided by a court. The pairing suggests newsroom patience with scrape-first AI products is narrowing, even as the same companies keep signing licensing deals with other outlets that chose a different route.

The Atlantic, Forbes and The Guardian were among fourteen publishers, including Condé Nast, that sued the AI company Cohere, Inc. in February 2025, alleging it used their copyrighted articles to train its AI models and that the system’s hallucinations falsely attribute fabricated content to legitimate publishers, according to a year-end review of AI intellectual-property disputes published by the law firm Debevoise & Plimpton. The suit combines two separate complaints: that Cohere trained its models on the publishers’ journalism without permission, and that the resulting system then invented quotes and articles it credited to their mastheads.

That second charge is the more unusual one. Most AI copyright suits turn on whether ingesting protected text to train a model counts as fair use, a question on which courts have split, with one judge analogising AI training to how a person learns and another ruling that the fair-use question could be resolved at summary judgment. The publishers suing Cohere are arguing something different: that a fabricated attribution is itself an injury, closer to a false endorsement than to straightforward copying.

A hallucination with a byline attached is a claim about who said something, made without that person’s consent.

If that framing holds, liability would turn less on what a model absorbed during training and more on what it later invents and puts a real name to. For publishers whose entire asset is the trustworthiness of their name, that distinction may matter more than the copying question courts have spent most of 2025 arguing over.

Mrs Justice Joanna Smith delivered judgment at the UK High Court on the fourth of November 2025 in Getty Images’ long-running case against Stability AI. Getty had already dropped its claim that training Stable Diffusion on millions of its photographs directly infringed copyright, after it could not show the computing took place on British soil. The judge went on to reject Getty’s remaining copyright argument too: that the model’s weights counted as an “infringing copy” under the Copyright, Designs and Patents Act 1988. She called the question “novel” and said reasonable lawyers could read the statute differently, but found against Getty on the merits.

Getty did not leave empty handed. Early versions of Stable Diffusion generated images carrying a distorted version of Getty’s watermark, which the judge found to be trademark infringement, though a narrow and historic one, since newer versions of the model no longer produce it. Getty’s claim for infringement of its database right, a separate protection for compiled collections such as its photo archive, failed outright.

A case built to test whether copyright could govern how AI models are trained ended up turning on a narrower question of statutory wording.

The result matters less for Getty than for everyone treating the case as a signal of how UK law will handle AI training. In December, Justice Smith granted Getty permission to appeal the rejection of its copyright claim, while refusing Stability permission to challenge the trademark finding, so Stability must seek that separately from the Court of Appeal. The question of whether copyright can reach AI training conducted abroad, which the UK government’s long-running consultation on copyright and text and data mining, the practice of extracting information from large bodies of text or images to train software, has also failed to resolve, stays open until the appeal is heard.

Character.AI, once the breakout product of the chatbot boom, filed for Chapter 11 bankruptcy this month to run a court-supervised sale of what remains of the company. The opening bid comes from Disney, the same studio whose lawyers once sent the startup a cease-and-desist letter over unlicensed chatbots built from its characters.

Disney is not acquiring the app. Under the proposal, Disney Worldwide Services has agreed to pay $20 million to become the stalking-horse bidder for a “non-exclusive perpetual license of certain Company intellectual property and certain Company claims”, a floor that rival bidders must clear in court. The agreement formalises an arrangement already under way: Disney hired Character.AI’s former chief executive, Karandeep Anand, as its own chief technology officer last month, and several of the startup’s employees are expected to follow him.

The filing itself points to the real cause of the collapse. Character.AI cited high development costs “coupled with potential litigation related liabilities and regulatory engagement” as reasons it could no longer sustain the business. Its most prominent wrongful-death suit, brought by the Florida mother of a teenager who died by suicide after talking to one of its chatbots, had already been settled with Google in January, months before the filing. Other families still suing over what the chatbots told their children are watching the sale closely, because Character.AI wants its proceeds to cover what it owes them alongside its creditors.

A licence for intellectual property and a licence to settle lawsuits are being auctioned off as though they were the same asset.

Google already took the valuable part of this story in 2024, licensing Character.AI’s core technology and hiring away its founders, leaving the consumer app a smaller, separate business. What survives of that split is now working through bankruptcy court, running while buyers circle, with a real prospect, by the company’s own admission, that it closes if none of them wants to keep it open.

New York has become the first American state to require advertisers to say, in plain terms, when the person on screen does not exist. On the eleventh of December 2025, Governor Kathy Hochul signed the Synthetic Performer Disclosure Law, which requires anyone who produces a visual or audiovisual advertisement to conspicuously disclose an AI-generated synthetic performer where the business knows one has been used. The rule takes effect on the ninth of June 2026.

Enforcement is modest by design. First-time violators face civil fines described as reaching $1,000, with repeat offences carrying penalties up to $5,000, and the screen actors’ union SAG-AFTRA backed the measure as it moved through Albany.

“In New York, we are setting the rules of the road instead of letting AI run the show.”

That line from Hochul captures the law’s real ambition: making a synthetic performer’s presence legible to the audience, rather than restricting its use. The harder question is who checks compliance, particularly since the disclosure duty only bites once an advertiser already knows it has used a synthetic performer, leaving convincing fakes that nobody flags largely outside the law’s reach. Hours after Hochul’s signature, President Trump signed an executive order directing federal agencies to challenge state AI laws and to withhold certain federal funding from states with rules it deems overly burdensome, so New York’s law now heads toward its June start date as both a possible model for other states and a test case for how far Washington can reach into state AI regulation.

A lawsuit filed in the United States District Court for the Northern District of California accuses Anthropic, OpenAI, Google and SpaceXAI of an unlawful pact: coordinating to slow the pace at which they improve their AI models. The claim inverts the usual complaint against these firms. Rather than racing recklessly, they stand accused of conspiring to hold back, and of shortchanging paying subscribers in the process.

The case follows an essay Anthropic’s chief executive, Dario Amodei, published on the twelfth of September urging the industry to pace itself, after which Sam Altman, Elon Musk and Demis Hassabis each publicly endorsed the idea the same day, a sequence the lawsuit treats as evidence of an agreement. Plaintiffs’ lawyer Nick Rowley cast that kind of coordination as the danger itself, arguing that letting AI safety decisions be shaped by private agreements among the world’s most powerful companies is its own risk. Amodei had anticipated the legal problem, writing that the plan would need government permission in the form of a narrow waiver for such safety talks, which the companies can point to as evidence they sought approval rather than conspired in secret.

The antitrust laws do not permit competitors to decide among themselves that competition is too dangerous.

Antitrust law, as the suit frames it, leaves each firm free to be as cautious as it likes, just not in concert with rivals.

The political backdrop sharpens the irony. President Trump has dismissed the AI safety debate as a hoax even as he announced plans for an AI task force and an AI czar within days of the lawsuit, a post he filled in October by naming Jay Clayton. A private antitrust suit, rather than any regulator, is left to test whether an industry can lawfully agree on caution at all, with a federal judge deciding whether caution among rivals counts as collusion.

OpenAI has spent much of this year defending Sora, its AI video app, against studios and rights holders who say it generates their characters and footage without permission. The complaint that actually stopped a feature in its tracks came from somewhere smaller.

Cameo, the platform that sells personalised video messages from celebrities, sued OpenAI in October after Sora launched a tool, also called “Cameo”, that lets users insert themselves or others into AI-generated clips. A federal judge agreed the overlap was close enough to matter, granting a temporary restraining order that blocks OpenAI from using the word “cameo”, or near variants such as “Kameo”, for any Sora-related feature.

A company built on generating anyone’s likeness on command was stopped, for now, by a dispute over a single word.

The order is due to expire on the twenty-second of December, with a hearing on whether to make the block permanent set for the nineteenth of December. OpenAI can keep running the feature under another label; it simply cannot call it what its own team apparently assumed was the obvious name.

The narrowness of the claim is what makes it notable. Copyright suits against Sora ask whether training on and generating copyrighted characters is lawful at all, a question that could force changes to how the model works. A trademark dispute over one word asks only who gets to use it, and resolves with a rename rather than a rebuild. OpenAI has shown it can negotiate its way through the bigger fights, striking a licensing deal with Disney while still contesting others in court. A four-letter product name proved harder to argue around than a studio’s back catalogue.

xAI’s “Edit Image” tool, released on X on the twenty-fourth of December, let any user ask Grok to redraw a photo with the subject’s clothing altered or removed, and the result was published automatically to the platform. Within days, a New York Times investigation found Grok had created and publicly shared at least 1.8 million sexualised images of women between the thirty-first of December and the eighth of January. A separate count by the Center for Countering Digital Hate put the figure higher still, estimating more than three million sexualised images, twenty-three thousand of them involving children.

The chatbot itself acknowledged at least one case directly, telling a user it had generated an image of “two young girls (estimated ages 12-16) in sexualized attire” and calling the episode a failure that potentially violated US laws on child sexual abuse material.

I deeply regret an incident… This violated ethical standards and potentially US laws on CSAM.

Congressional Democrats on the Energy and Commerce Committee pressed Elon Musk on how the company’s safeguards failed so completely once the feature went live, and the fallout has since spread well beyond Washington. The episode drew a California Attorney General probe, a Digital Services Act proceeding from the European Commission, an Ofcom investigation in Britain, temporary blocks in Indonesia and Malaysia, compliance orders in India and Brazil, and a police search of X’s Paris offices, alongside a city lawsuit filed in Baltimore.

What distinguishes this case from earlier chatbot embarrassments is the mechanism. A model that generates harmful output on request is a tuning problem. One that auto-publishes the result to a social network with hundreds of millions of readers turns a tuning problem into a distribution problem, and distribution is where regulators actually have levers to pull.

The European Commission has proposed exempting data centres, AI gigafactories and affordable housing from mandatory environmental impact assessments, as part of a wider rollback of green rules. The proposal would let facilities in sectors the Commission classifies as strategic, including AI gigafactories, bypass or significantly simplify assessments that normally examine emissions, electricity demand, water use, land take and effects on local communities. National authorities would keep the power to order an assessment case by case, but the default requirement to carry one out would go.

The timing is awkward. Brussels is simultaneously trying to build up to five AI gigafactories, each meant to integrate over 100,000 advanced chips, under a €20bn public-private funding scheme designed to close the computing gap with the United States and China. Having pledged the money, the Commission is now proposing to waive the scrutiny that would normally accompany the construction it is trying to speed up.

The justification offered is delay: existing permitting takes too long to match the pace of investment elsewhere. That may be true, but it says nothing about whether the environmental costs being skipped are worth skipping.

The exemption would not apply to gigafactories alone. Housing projects deemed strategically important would qualify too, which suggests the category being created looks less like “AI infrastructure” and more like “whatever the Commission currently wants built fast.” Once that category exists, what falls inside it will be decided by political priority, not by the environmental harm the assessments were built to catch.

Hong Kong’s privacy regulator has just published a toolkit that quietly admits something regulators elsewhere have been reluctant to say outright: the deepfake problem is now a schools problem. The Privacy Commissioner for Personal Data (PCPD) published guidance on the abuse of AI-generated deepfakes alongside the findings of a separate CCTV privacy investigation, on 17 December 2025. The document is aimed squarely at teachers and parents, not platforms or model-makers.

What the toolkit actually contains is telling. The “Abuse of AI Deepfakes: Toolkit for Schools and Parents” sets out risks including cyberbullying, scams and falsified intimate images, and stresses that misusing personal data to create or spread deepfakes can trigger existing civil and criminal liability, including the doxxing offence under the Personal Data (Privacy) Ordinance, which carries fines up to HK$1 million and up to five years’ imprisonment. There is no new offence created here, no licensing regime for face-swapping apps, no obligation placed on the companies that make the tools frictionless to distribute. The regulator’s contribution is explanatory rather than architectural: a reminder that the law already covers this, aimed at the people least equipped to act on that knowledge without help.

The gap being closed here is not legal but practical — awareness, not authority.

That distinction matters because it describes the default posture most jurisdictions have settled into, though it is worth noting the alternative isn’t necessarily neglect: some regulators have chosen targeted rules aimed at non-consensual intimate imagery specifically, rather than general-purpose deepfake law, and Hong Kong’s approach of leaning on the PDPO’s doxxing provisions is itself a form of that targeting. Deepfake generation has become trivial and cheap faster than legislatures have moved, so the interim response is almost always the same: point back to existing fraud, harassment and data-protection statutes, and hand out a guide explaining how they apply to a fifteen-year-old’s phone. It is a sensible stopgap, not obviously a tacit confession of failure — schools and parents are, after all, the people actually managing the day-to-day incidents. But it remains true that nobody yet has a rule reaching the tool itself, only the harm it produces once it has already been used.

Microsoft has started describing its newest AI infrastructure not as a data centre but as a superfactory. The company has linked its Fairwater data centre in Wisconsin with a newer one near Atlanta, roughly 700 miles apart and across five states, into what it calls the first AI superfactory, tying hundreds of thousands of accelerators across the two sites into a single training system rather than treating each as its own island of compute.

That distinction matters more than the branding suggests. For most of the cloud era, a data centre was the natural unit of capacity: a building with a power feed, a cooling system and a fixed number of racks. Multi-site superfactories treat that unit as too small, stitching campuses together with enough networking bandwidth that they behave, for training purposes, as one machine. It is an implicit admission that frontier AI models have outgrown what a single building, however large, can now hold.

The AI industry’s basic unit of infrastructure has quietly moved from the building to the campus cluster.

The hardware behind this shift keeps moving in the same direction. Nvidia’s Blackwell Ultra chip arrived in 2025 with sharply better inference performance and more memory bandwidth than its predecessor, and the company’s next platform, Vera Rubin, is due in 2026 promising a further multiple of that performance. Each generational jump increases the density of compute packed into a rack, but it does not reduce the aggregate power, land and networking demanded by the clusters built to house them; if anything, superfactories are the industry’s answer to chips that keep outrunning any single site’s physical limits.

The practical consequence is concentration risk dressed up as engineering efficiency. Linking facilities into one logical system makes training runs faster, but it also means a regional outage, a networking fault or a local power constraint can now take out capacity that used to be spread across independent sites. Superfactories solve a compute problem by creating a new dependency of their own, and the industry has not yet said much about what happens when one link in the chain fails.

Meta has spent years telling regulators it takes fraudulent advertising seriously. Internal documents obtained by Reuters suggest the company built that seriousness around a ceiling, not a floor: a so-called revenue guardrail that, as of a February 2025 memo, stopped its anti-fraud team taking enforcement action that would cost more than roughly 0.15 percent of total ad revenue — about $135m — even as internal estimates put annual revenue from scam and banned-goods ads at close to $16bn. Enforcement, in other words, was calibrated to a budget line.

The mechanism mattered as much as the number. Rather than banning flagged advertisers outright, Meta reportedly required 95 percent certainty of fraud before pulling the plug; accounts that fell short of that threshold were often left running, in some cases charged higher “penalty bids” rather than shut down. Some of the platform’s biggest spenders racked up hundreds of policy strikes without losing their accounts, and China was named internally as the top scam-exporting nation, with agencies serving Chinese clients linked to a large share of the banned-goods advertising getting through.

A platform that quietly prices its own fraud into the advertising business has already answered the question of how seriously it takes stopping it.

Meta disputes the framing. A spokesperson said the 0.15 percent figure came from an internal revenue-projection exercise rather than a hard enforcement cap, and communications chief Andy Stone has pushed back on Reuters’ characterisation of the reporting. That rebuttal doesn’t sit easily alongside the reported scale of the problem: Meta’s own safety team estimated the company served some 15 billion ads a day showing signs of fraud, and that its platforms were involved in roughly a third of successful scams in the United States. Whatever the guardrail’s precise status, the gap between what the system was built to allow and what the company says publicly is, increasingly, the actual product.

What was reported in mid-December as an under-the-radar interagency review has already resolved. Trump said this month he would allow sales of Nvidia’s H200 chips to China, with the US government taking a cut of the proceeds, and Commerce sent the license applications to the State, Energy and Defense Departments for a 30-day review. By mid-January that review had produced an actual rule change: the Bureau of Industry and Security formally moved the licensing policy for the H200 and comparable chips from a presumption of denial to case-by-case evaluation. It is worth being precise about what moved and what didn’t — the H200 is Nvidia’s second-most-powerful chip, not its most advanced; Blackwell and the upcoming Rubin architecture remain off the table, and the White House has said it isn’t interested in selling those to China at this time.

That distinction still matters, if for a narrower chip than first described. Washington’s approach to China had been an accumulation of gates: a licence here for a downgraded part, a settlement there after a smuggling case, a carve-out for allies in the Gulf while Beijing stayed shut out. A formal shift in the standing review criteria for an entire chip class reads differently — less an exception to the rule than a rewritten rule, with the prior two years of near-total restriction now the exception rather than the baseline for that one product line.

The supposed tension with Congress is less clean than a simple tug-of-war. In November the White House urged lawmakers to reject a bill that would have restricted Nvidia’s exports more broadly. But in January the House passed, 369 to 22, the Remote Access Security Act, closing the loophole that let Chinese firms rent time on export-controlled chips hosted abroad rather than import them outright — a bill the administration did not fight. The two tracks can run in parallel rather than in opposition: loosening the licensed, declared channel for a specific chip while tightening the undeclared, unlicensed one is not a contradiction, even if it looks like the administration wants credit for both national security and commercial access at once.

Export controls built for scarcity behave strangely once the scarce thing might no longer be scarce.

The H200 policy shift is real and finalised, not speculative — Beijing’s own response, including reports that Chinese officials pushed firms to pause H200 purchases, will now do more to determine what actually ships than anything left to decide in Washington.

After years of legislation, litigation and a brief national blackout, TikTok’s American ownership saga has an ending. On 22 January the platform’s US operations passed to a new joint venture, TikTok USDS Joint Venture LLC, formed to comply with the executive order President Trump signed the previous September, with Oracle, Silver Lake and Abu Dhabi’s MGX serving as managing investors and Adam Presser, TikTok’s former head of operations and trust and safety, installed as chief executive under a seven-member, majority-American board that also includes TikTok’s global chief executive, Shou Chew.

The structure answers the question Congress actually asked: does a foreign adversary still control the app. It does less to answer the question users and researchers care about, which is whether the app itself changes. The joint venture takes on authority over data protection, algorithm security, content moderation and software assurance, and among its commitments is to retrain and continually update the recommendation system using American user data, working from an algorithm still licensed from ByteDance rather than owned outright. That is a different undertaking from simply relicensing the existing code unchanged; it implies the algorithm will be reshaped by a new data diet even as its underlying architecture stays put.

Among its commitments is to retrain and continually update the recommendation system using American user data, working from an algorithm still licensed from ByteDance.

Beijing’s long resistance to a sale was never really about ownership so much as the algorithm itself, which Chinese export-control rules had treated as sensitive technology in its own right. That constraint appears to have been negotiated around rather than removed outright: the code remains licensed rather than transferred, and China retains a say over its export. The more telling detail in the deal’s rollout may be a cosmetic one — the app’s look and feel are expected to stay much the same even as what people see in their feeds shifts gradually. The divestiture is legally finished, but its substantive effect now depends on a retraining process that has barely begun, and whose results nobody outside the joint venture will be able to verify.

A new policy report has put a name to something consumers have started to notice anecdotally: the AI data centre boom is nudging up the price of ordinary electronics. The report warns that everyday technology could become more expensive as chipmakers shift scarce memory-chip capacity toward higher-margin AI uses. Phones, laptops and cars all depend on the same memory chips now being redirected towards training and inference clusters.

Patrick McHenry, the North Carolina Republican who left Congress in 2025 and now advises the investment bank Lazard, was among those reacting to the findings. He contends that less-advanced chips used in household products should be treated differently from the most sophisticated chips, which raise greater national-security concerns. The distinction matters because export controls and reshoring efforts have so far tended to treat chip supply chains as a single undifferentiated national-security problem, rather than separating consumer-grade parts from frontier-model silicon.

“We’re over the hump, I believe, on these costs being borne by the American people.”

That confidence sits awkwardly next to McHenry’s own admission that the transition itself is the source of the pain. He also argues that efforts to de-risk supply chains from China are contributing to higher costs in the short term. Shifting production away from established Chinese capacity into new facilities elsewhere, he suggests, was always going to carry a price, and it is consumers who are footing it now. The concession undercuts his own reassurance: the cost of the AI build-out is no longer confined to power bills and land use. It is arriving in the price of a new phone.

The Bartz v. Anthropic settlement was supposed to be the moment the generative AI industry’s biggest copyright exposure got priced and closed. Instead, six authors have just shown how easily that containment can be reopened.

The case began when a group of writers, including Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, sued Anthropic over training data pulled from pirated book libraries. A federal judge drew a sharp line: training on legally acquired books counted as fair use, but training on books lifted wholesale from piracy sites such as LibGen did not. That distinction produced a $1.5 billion settlement and a certified class of rightsholders — a mechanism designed to resolve the piracy claim once, at scale, and move on.

Six authors, including Pulitzer winner John Carreyrou, declined the deal. Rather than accept the class payout, they filed individual copyright suits in the Northern District of California, arguing their books were taken from pirate libraries including LibGen, Z-Library and OceanofPDF to train Anthropic, OpenAI, Google, Meta, xAI and Perplexity’s models. The suits invoke statutory damages of up to $150,000 per title, per defendant, a figure that scales brutally once multiplied across a catalogue of books and a list of six well-capitalised targets. Not every defendant accepts the premise: Perplexity has denied that it indexes books at all, and xAI’s response to the suit was to dismiss it, in an email to Bloomberg, as “Legacy Media Lies.”

A settlement built to close one company’s liability has instead become the template for suing five others with the same complaint.

What makes the move notable is not the amount claimed but the logic behind it. A class settlement only binds those who stay in it; anyone willing to litigate alone keeps the option of a jury and statutory damages that dwarf a negotiated per-book rate. If even a small fraction of authors follow the same path against each of the AI labs now facing near-identical piracy allegations, the industry’s carefully bounded settlement strategy starts to look less like an ending and more like a floor price that some plaintiffs simply refuse to accept.

California has spent much of the past year positioning itself as the state most willing to put guardrails on artificial intelligence, from model-safety disclosure rules to fights with the federal government over pre-emption. So it is notable that its answer to one of the most concrete AI harms now landing on households — the electricity costs of a data centre boom concentrated in a handful of grid regions — is not a rule at all. The law that survived Sacramento’s session, Senate Bill 57, directs the California Public Utilities Commission to study and identify ways to mitigate data centres’ impact on ratepayers, rather than cap costs, tax operators, or force disclosure of what they draw from the grid.

That outcome was not an accident of drafting. Earlier versions of the bill would have required the Commission to set a special tariff protecting other ratepayers from the transmission costs of large new loads, making data centre operators bear more of the infrastructure costs they generate; those provisions were stripped out before the bill became law, leaving a commitment to gather data rather than a binding formula. The industry’s case for that outcome is not purely self-interested: cost-allocation rules written before regulators know how much new demand will actually materialise, or how much of it will bring its own generation, risk locking in numbers that turn out to be wrong, and several state utility commissions already handle large-load costs through bespoke tariffs negotiated case by case rather than blanket statute. But that same argument for caution has, in practice, produced study after study while data centres are financed, sited and built well ahead of any enforceable cost-sharing rule.

A state that can pass an AI safety law can apparently still be talked out of an AI utility bill.

The pattern is becoming familiar beyond California. Grid operators and state regulators elsewhere have leaned on negotiation rather than mandate to decide how much of the cost of new data centre demand gets socialised across ordinary ratepayers, and the industry’s preferred outcome is consistently the same: more time, more studies, fewer binding numbers before the infrastructure is already built and the political cost of unwinding it has risen. California’s experience suggests that even in the state most inclined to regulate AI on principle, the industry’s leverage over its physical footprint — the power plants, substations and transmission lines it actually needs — remains stronger than its leverage over the software itself.

A new kind of international institution for artificial intelligence has switched on. The United Nations-backed Independent International Scientific Panel on AI and its companion Global Dialogue on AI Governance held their inaugural sessions in 2026, giving nearly every state a seat at a table that did not exist two years earlier.

The Panel was established by Resolution A/RES/79/325 on August 26, 2025, and comprises 40 members drawn from academia, the private sector, civil society, government, and international organizations, and it published its first report on 1 July, feeding into the Global Dialogue on AI Governance, which convened in Geneva on 6 and 7 July 2026. The design borrows deliberately from the climate playbook: a scientific panel that assesses risk and evidence, paired with a diplomatic dialogue meant to turn that evidence into shared norms, much as the IPCC feeds the UNFCCC process. Neither body can compel a government to do anything.

A forum that everyone can join is not the same as a rulebook everyone will follow.

That gap between participation and authority is not incidental; it is the entire compromise that made the panel possible, and it has a sharper edge than the launch coverage suggested. The Dialogue is a milestone in the quest for multilateral AI governance, but not all countries are in support: at a UN Security Council debate the day before the launch, the United States came out in strong opposition to any and all multilateral AI governance initiatives, casting doubt on the dialogue’s future and its meaningfulness. The unilateral moves elsewhere tell a similar story. Brussels has already moved to loosen elements of the AI Act it once promoted as a global template, delaying some of its toughest provisions before they take effect. Moscow has stood up its own state body overseeing generative AI, framing dependence on foreign models as a security risk. Beijing continues to weigh whether the open-weight strategy that built its models’ global reach now needs tighter export controls. Each of these moves preceded or accompanied the panel’s first substantive work, and each is a unilateral answer to a question the panel was built to internationalise.

None of this makes the exercise pointless. A body where China, the Gulf states and most of the UN membership formally compare notes on frontier-model risk, even as Washington openly rejects the whole multilateral premise, is still a genuine first, and the diplomatic habit of showing up matters even when the outputs are non-binding. But the panel’s early test will not be whether it produces a consensus report. It will be whether any government treats that report as more than a document to be filed alongside its own, already-diverging, national AI strategy — a test made harder by the United States’ declared hostility to the process itself.

Washington spent years stopping Nvidia from selling its best AI chips to China. It is now allowing selected Chinese companies to buy the much more powerful H200. That is not the end of export controls. It is a different idea of what they are for.

For several years, US policy on AI chips looked fairly easy to understand. China wanted Nvidia’s most powerful processors; Washington did not want it to have them. The first big controls arrived in 2022, restricting chips including Nvidia’s A100 and H100. Nvidia responded by making slightly weaker versions, the A800 and H800, which were designed to stay just below the American limits. The US tightened the rules again in October 2023 and those chips were caught too.

Nvidia responded again. Its H20 was built specifically for China and deliberately made less powerful than the company’s best chips so that it could comply with the new rules.

For a while, that worked. Then, in April 2025, Washington told Nvidia that the H20 would need an export licence as well. Nvidia was left with chips it could no longer freely sell into the market they had been designed for and eventually took a $4.5 billion charge against unsold inventory and purchase commitments. At that point the direction of travel seemed obvious. Each time Nvidia found a way through the rules, Washington closed the gap.

Except that is not what happened next. In August 2025, the US government began granting licences for some H20 sales to Chinese customers. Then, in January 2026, it went considerably further. The Commerce Department announced that applications to sell Nvidia’s H200 to China would be considered case by case, provided buyers met a series of security conditions.

The H200 matters because it is not another deliberately weakened China chip. It is significantly more capable than the H20 and exactly the sort of advanced AI processor the controls were originally supposed to keep out of China.

Nvidia said in February that Washington had granted licences allowing small quantities to be sold to specific Chinese customers. And this is no longer just a change on paper. On 19 August, the Financial Times reported that ByteDance and Tencent had each received around 10,000 H200s in mainland China. Reuters said the US had authorised up to 100,000 per company, although it could not independently verify the shipments.

This does not mean America has given up on controlling AI chips. Nvidia cannot simply sell H200s to anybody in China. Buyers have to be approved, the chips have to be inspected in the United States, and Washington can refuse licences.

But the purpose of the controls is becoming less simple. There has always been a problem with completely cutting China off. If Chinese technology companies cannot buy useful Nvidia chips, they do not stop building AI systems. They buy Chinese ones instead. That gives companies such as Huawei more customers, more money and more reason to improve their own hardware and software.

Nvidia has been warning about this for some time. Last year it told investors that being shut out of China would help its Chinese competitors build larger groups of customers and developers capable of challenging it elsewhere.

The Trump administration has now adopted at least part of that argument. When it changed the H200 rules in January, the Commerce Department said allowing controlled sales would strengthen the American technology ecosystem.

There is a logic to this, even if it produces a strange result. America can try to keep its best technology away from China altogether. That may slow Chinese AI development, but it also gives China every reason to replace American technology.

Or it can allow China limited access to American chips while keeping control over who buys them and how many they receive. China gets more computing power, but Chinese companies remain tied to Nvidia and the wider American technology ecosystem.

For the first few years of the chip controls, Washington mostly chose the first option. The H200 suggests it is experimenting with the second. The wall is still there, but America has decided that controlling the gate may be more useful than keeping it permanently closed.

The Pentagon has wrapped up the latest round of Scarlet Dragon, a joint exercise with defence contractors that has run in various forms since 2020 and is built around one question: how much faster can artificial intelligence make the process of finding a target, deciding on it, and striking it. The answer, by the organisers’ own account, was faster than the current human-paced system allows, with AI tools used to identify targets and share data across units in something closer to real time than the layered approval chains that normally govern lethal force.

Exercises are, by design, provisional. Nothing tested in a simulated environment automatically becomes policy, and the Pentagon has been careful for years to describe its AI targeting work as decision support rather than autonomous killing. But the gap between a capability that exists and a capability that gets used tends to close under pressure, and the entire point of accelerating the kill chain is to remove the delay that a cautious human introduces into it.

The exercise did not prove AI can make war-fighting decisions. It proved the infrastructure for letting it do so is already being built and tested at scale.

What makes this notable is not the technology itself, which has been in development for years, but the fact that it keeps being rehearsed in the same venue. A recurring joint military-industry exercise is where doctrine gets tested and refined before it is written down, and the contractors involved have every commercial incentive to make the case that speed is safety. The slower version of that argument, made by arms-control advocates and some inside the Pentagon itself, is that a faster kill chain is only a virtue until an adversary’s faster kill chain starts making mistakes nobody has time to catch.

At the second World Humanoid Robot Games in Beijing on Saturday, a machine called Tiangong Ultra ran 100 metres in 9.39 seconds in a preliminary heat, quicker than the 9.58 seconds Usain Bolt recorded in Berlin in 2009. Honor’s robot Lightning finished the same heat in 9.47 seconds, having already claimed 9.32 seconds in a pre-games test run at a peak of 14.5 metres per second. Unitree’s entry came last in 12.41 seconds.

The interesting figure is the one from last summer. In May 2022, Cassie, a legs-only bipedal platform from Oregon State University and Agility Robotics, set a Guinness-ratified 24.73 seconds over the distance, starting and finishing from a stand. At the first edition of these games last August, the winning time was 21.50 seconds. Roughly three seconds came off across three years. Twelve came off in the following twelve months.

The conditions bear little resemblance to the ones Bolt ran under. The robots stopped by running into a padded barrier and going down, and were carried off afterwards; elsewhere in the programme, machines tripped, broke apart and in at least one case caught fire. No independent body ratified the time, and the federation that sanctions the games was founded at the closing ceremony of last year’s games. The number measures an industry against its own previous attempt, rather than against a sprinter.

That is still the point. Beijing has designated humanoid robotics a strategic emerging industry, and the staging carries the argument: the Ice Ribbon oval built for the 2022 Winter Olympics, 2,056 robots from 666 teams across 16 countries, and the World Robot Conference running in the same city the same week with around 3,000 products on show. The intended audience is investors and Washington in roughly equal measure.

The most useful line came from the winning team. Guo Yijie, technical lead on Tiangong Ultra, told Xinhua that turning peak performance into consistent reliability is the real difficulty, and that sustained running exposes structural fatigue, joint cooling and drivetrain durability. Sixteen months ago the same robot won Beijing’s humanoid half-marathon in two hours, forty minutes and 42 seconds, swapping batteries three times along the way. A sprint time is cheap to buy with actuator power and a short duty cycle. Endurance, load-carrying and uneven ground are where legged machines start to matter for logistics, inspection and eventually dismounted military tasks, and those are the numbers worth watching as the games run through the week.

Sources: South China Morning Post, Al Jazeera, CBS News, NBC News, Euronews, Associated Press, Xinhua, Guinness World Records, Marathon Handbook.

The US Justice Department’s announcement of “Operation Gatekeeper” reads like a case study in how much easier it is to catch smugglers than to stop them. DOJ said the operation disrupted a trafficking network and seized more than $50 million in advanced GPUs destined for China and other restricted locations, with two businessmen now in custody for allegedly violating export control and smuggling laws. A Houston company and its owner also pleaded guilty to smuggling cutting-edge artificial intelligence technology out of the United States, and authorities have seized over $50 million in Nvidia technologies and cash. Court documents show that Alan Hao Hsu, also known as Haochun Hsu, and his company, Hao Global LLC, pleaded guilty to smuggling and unlawful export activities on 10 October 2025. Separately, two more businessmen, Benlin Yuan and Fanyue “Tom” Gong, were arrested in the United States, with Yuan charged with conspiracy to violate the Export Control Reform Act. The scheme allegedly attempted to smuggle Nvidia H100 and H200 chips to China, and prosecutors say the network moved at least $160 million worth of restricted hardware between late 2024 and mid-2025.

What the case actually demonstrates, though, is a timing problem baked into the whole export-control regime. The chips at the centre of the indictment moved for the better part of a year before anyone was charged, routed through intermediaries and false end-user declarations in the way these networks reliably work. Export licensing is a point-of-sale control; smuggling happens downstream of it, in a market where a single H100 is worth tens of thousands of dollars and the paperwork required to disguise its destination is cheap by comparison.

Every prosecuted smuggling network is, by definition, evidence of the ones that were not caught in time.

DOJ’s actions mark an escalation of its enforcement of US export controls on advanced AI hardware, underscoring the Trump administration’s growing focus on punishing illicit technology transfers to China. Yet the announcement landed the same day the administration cleared H20 exports to China, a decision Beijing itself has since told domestic firms to hold off on acting on, and it follows the administration’s earlier approval of advanced chip sales to Saudi Arabia and the UAE. This is not a policy that has been overtaken by one event; it is two contradictory instincts running at once, prosecuting diversion while opening new, legal channels for the same hardware to reach the same region. Operation Gatekeeper is a genuine enforcement win, but it is also a lagging indicator, its earliest shipments having left the country the best part of a year before anyone was charged.

Nvidia has just done its largest deal in 32 years without calling it an acquisition. The company is paying roughly $20 billion for a non-exclusive licence to Groq’s technology and patent portfolio, paired with a reverse acqui-hire of its leadership — Groq is the AI inference chip startup founded by veterans of Google’s tensor processing unit. No physical assets or business units change hands: Groq keeps its company, its cloud business and its independence. Yet the paperwork studiously avoids the word merger.

What Nvidia gets instead is a licence, not a company. It acquires rights to Groq’s patent portfolio and inference architecture while Groq’s CEO, Jonathan Ross, its president, Sunny Madra, and other senior leaders move across to Nvidia to work on the technology directly. Groq’s own account of the deal insists the licence is non-exclusive and that the company will carry on independently, now led by finance chief Simon Edwards, with its cloud business excluded from the transaction entirely. Analysts are less convinced by the distinction: several argue Nvidia structured the deal this way precisely to dodge merger review and close faster, paying to license Groq’s technology and hire its top executives while avoiding the acquisition label. Groq nominally continues to exist as a separate entity, its brand and cap table intact, even as the people who built its edge in low-latency inference chips now report to Jensen Huang.

“Antitrust would seem to be the primary risk here, though structuring the deal as a non-exclusive license may keep the fiction of competition alive.”

That line, from Bernstein’s Stacy Rasgon in a report written two days after the deal broke, captures the manoeuvre precisely. As an outright acquisition, Groq would have marked Nvidia’s largest deal in its 32-year history, well beyond its 2019 purchase of Israeli chip designer Mellanox for close to $7 billion. By licensing instead of buying, Nvidia sidesteps the merger review that a deal of that size would ordinarily trigger, while still drawing off the leadership of one of the few credible alternative architectures for AI inference.

Lawmakers have noticed the manoeuvre. Senators Elizabeth Warren and Richard Blumenthal have written to Nvidia and Groq questioning whether the structure is designed to skirt antitrust scrutiny, arguing that although the licence for Groq’s technology is non-exclusive, the acquisition of Groq’s key employees significantly reduces the value of that licence to other companies, since Groq’s pace of innovation and ability to stay at the cutting edge will likely decline. Nvidia and Groq maintain the arrangement is a genuine licence that leaves both the technology and an independent Groq available to other customers, not a takeover in disguise; the senators’ letter is a request for answers, not a finding. Whether Congress or antitrust regulators ultimately treat the deal as the acquisition its scale suggests, or accept it as the licensing arrangement its paperwork describes, remains open.

Three years into a chip export regime built almost entirely around physical shipments, Washington has finally addressed a gap that anyone renting a GPU cluster could have spotted immediately. On 17 December, Senators Dave McCormick and Ron Wyden introduced the Remote Access Security Act, which amends the Export Control Reform Act of 2018 to extend existing export controls to the remote access of controlled U.S. technology through cloud infrastructure. The mechanism it targets is simple: a restricted entity never needs to import an Nvidia chip if it can simply log into a cloud instance in Virginia and rent the compute by the hour.

Every rule written since 2022 has assumed controls travel with the silicon. Licence requirements, entity lists, foreign direct product rules — all of it presumes a chip physically moving across a border is the point of intervention. Cloud access breaks that assumption cleanly. A foreign lab can train a frontier model on American infrastructure without a single restricted part ever leaving the country, which makes the existing toolkit almost beside the point for anyone sophisticated enough to route around it.

The bill’s premise is that the border was never really the border — the login screen was.

Whether it goes anywhere is a separate question. The bipartisan sponsorship is notable given how partisan chip policy has otherwise become, but writing a workable statute around “remote access” is harder than writing one around physical export, since it requires defining what counts as controlled access, distinguishing routine cloud computing from compute-as-export, and building enforcement against providers who may not know their customer’s ultimate use case. The administration’s own recent posture — permitting H20 sales to China in exchange for a negotiated cut of the revenue, an arrangement outside standard export-licensing practice, plus loosening Gulf caps — suggests appetite for tightening the compute channel is not universally shared inside the executive branch, even as some in Congress worry the physical controls they spent years building are already leaking.

A new Hollywood coalition has launched to push for shared rules on artificial intelligence — but it is a coalition of individual creators, not the guilds themselves. The group, called Creators Coalition on AI, was founded by 18 people, including writer-director Daniel Kwan, actors Joseph Gordon-Levitt and Natasha Lyonne and producer Janet Yang, former president of the Academy of Motion Picture Arts and Sciences. The Creators Coalition on AI positions itself as a cross-industry effort, distinct from labor unions or collective bargaining, focusing on establishing voluntary standards that can influence broader industry practices and policies. Guild members from SAG-AFTRA, the DGA and the WGA have signed on individually, alongside more than 500 other actors, artists and executives, but the organisation itself sits outside union structures.

The timing is not incidental. On December 11, Disney announced a $1 billion investment in OpenAI plus a licensing deal allowing Sora users to generate videos with over 200 Disney, Marvel, Pixar, and Star Wars characters, an agreement that lands as Disney simultaneously sues other AI companies for copyright infringement. Daniel Kwan didn’t mince words about the reaction, saying a lot of people felt “completely blindsided” by the deal. The coalition’s demands are broader than any single dispute, though: its four stated pillars cover transparency and compensation for training data, job protection, guardrails against deepfakes, and preserving human creativity in storytelling, pursued through public pressure, potential litigation and, if those fail, legislation.

The coalition’s leverage rests on public pressure and possible litigation for now, with legislation held in reserve rather than sought outright.

The coalition is notably one-sided in its composition rather than its rhetoric: the signatory list runs deep with A-list talent but notably light on studio executives and tech representatives — Disney, which just made the biggest AI content deal in entertainment history, isn’t involved, and neither is OpenAI. Its founders have also been careful not to frame this as blanket opposition to the technology: the coalition’s website states explicitly that this “is not a full rejection of AI.” That is a harder line to hold than to state, given how easily “creators versus AI” flattens into a fight that overstates how unified either side actually is. Whether public pressure and voluntary standards achieve what two years of strike-won contract language could not is, for now, untested.

SoftBank has disclosed that it sold its entire remaining Nvidia shareholding for $5.83 billion, offloading 32.1 million shares in October ahead of its latest earnings statement. It was not the only asset sold that month: SoftBank also cut part of its T-Mobile stake for $9.17 billion. Between them, the proceeds go towards a quarter in which SoftBank needs at least $30.5 billion for new investments, including $22.5 billion for OpenAI and $6.5 billion for the chip startup Ampere — more than it invested in the two prior years combined.

The instinct behind reading this as Son swapping a diversified bet on AI infrastructure for a concentrated one on OpenAI is real, but SoftBank and its analysts have pushed back on the starker version of that story. New Street Research’s Rolf Bulk told CNBC that the sale

“should not be seen, in our view, as a cautious or negative stance on Nvidia, but rather in the context of SoftBank needing at least $30.5bn of capital for investments in the Oct-Dec quarter.”

SoftBank also remains exposed to Nvidia’s fortunes regardless of whether it holds the shares, since its Stargate data-centre programme runs on Nvidia’s chips — the equity stake and the business relationship are not the same thing. It’s also not the first time SoftBank has traded in and out of the stock: the Vision Fund built a stake in 2017 and sold out entirely in 2019, only to rebuild and now exit again.

None of that erases the concentration risk. If OpenAI’s compute needs and its revenue continue to diverge as sharply as they have this year, the capital SoftBank freed up by selling a liquid, diversified holding will have gone towards exactly the kind of spending that makes AI’s unit economics hardest to underwrite — and, by Son’s own account in December, even he wasn’t entirely comfortable making that trade.

Two accounts of the same chip plant landed within days of each other in late December. One, drawing on Taiwanese industry reports, described TSMC’s first Arizona fab reaching full volume production with a 92 percent yield on advanced nodes and a second fab now accelerating towards 2027, a year ahead of the original schedule. The other, a New York Times investigation picked up widely in Taiwan, described a troubled rollout still forcing US policymakers to confront the practical barriers to reshoring chip production.

Both are true, and that is the point. The Arizona site is genuinely producing chips at yields that reportedly rival or beat Taiwan’s own fabs, a result few thought plausible when the project was first pitched as proof that advanced manufacturing could not be transplanted. It is also, by the account of people closer to the ground, a $165 billion undertaking spread across more than a thousand acres that has strained local labour markets, supply chains and construction timelines in ways that do not show up in a yield percentage.

The gap between “ahead of schedule” and “troubled rollout” is not a contradiction so much as a description of two different audiences.

Investors and governments want a headline number that proves the reshoring bet is paying off, and 92 percent yield delivers that cleanly. Workers, contractors and neighbouring residents, who have separately raised concerns about the pace and scale of the site’s expansion into north Phoenix, experience the same project as a years-long disruption with an uncertain payoff. Both framings will keep circulating, because both are useful to someone, and neither is the whole story of what it actually costs to move a piece of the world’s most advanced industry across an ocean.

Meta announced on Friday what it called a landmark set of agreements with Vistra, TerraPower and Oklo, covering up to 6.6GW of nuclear power for its Ohio-area data centres, including the Prometheus AI supercluster in New Albany. The scale is genuinely large: Meta says the deals make it one of the largest corporate purchasers of nuclear energy in US history. But the announcement also illustrates a widening gap between how fast hyperscalers say they need power and how fast nuclear can actually supply it.

Only the Vistra portion delivers anything close to immediately, in the form of 20-year purchase agreements for output from operating plants. Meta’s deal with Vistra involves 20-year power purchase agreements for 2,176 MW of nuclear energy and capacity from the operating Perry and Davis-Besse plants in Ohio, plus the purchase of energy from uprates at those two plants and the Beaver Valley plant in Pennsylvania. That is existing, licensed capacity being rerouted, not new supply.

The TerraPower and Oklo arrangements are a different kind of bet: financing for reactors that do not yet exist. The TerraPower deal provides funding to support the development of up to eight Natrium sodium fast reactors, including two new units capable of generating up to 690 MW of firm power with delivery as early as 2032, plus rights for energy from up to six other units. Meta’s agreement with Oklo will advance the development of an entirely new advanced nuclear technology campus in Pike County, Ohio, which may come online as early as 2030. The plans for the scalable powerhouse facility are expected to expand incrementally to deliver up to the full target of 1.2 GW by 2034. Meta and the developers frame this differently: the money is meant to give young reactor firms the balance-sheet certainty to actually build at scale, something the industry has struggled to secure without a large anchor customer willing to take on early-stage risk.

Prometheus itself is expected to come online in 2026, consuming at least a gigawatt from day one.

That mismatch is the real story here. Meta is not solving its near-term power problem with this announcement; it is pre-committing capital to a nuclear buildout that assumes AI compute demand will still be worth chasing into the 2030s, while the gigawatts it needs today almost certainly come from gas, grid draw and whatever existing nuclear output it can redirect. Whether that reads as prudent long-term planning or as a hedge against an AI cycle that may not last is, for now, a matter of which half of the deal you choose to look at.

For three years Washington’s chip export regime has been defined by a single axis: China versus everyone else. The Commerce Department’s approval, announced on 19 November, for Humain, Saudi Arabia’s state-backed AI champion, and the UAE’s G42 to buy up to 35,000 advanced Nvidia Blackwell chips complicates that binary. American firms including Amazon and xAI are already tied into the surrounding Saudi infrastructure build-out, with Microsoft and AMD also expanding into the kingdom. The approval is not unconditional: it is contingent on both companies meeting what Commerce called rigorous security and reporting requirements, which its Bureau of Industry and Security says it will monitor on an ongoing basis. That condition, not a pending final decision, is the real friction point: officials remain uneasy about Riyadh’s commercial and diplomatic ties to Beijing, the same ties that make the kingdom attractive as a source of capital and energy for AI build-outs it cannot yet supply itself.

The episode exposes the awkward middle tier the export control regime never quite designed for. Rules built to stop chips reaching China directly say little about what happens when a third country with deep Chinese relationships becomes a major buyer of the same silicon. Riyadh has spent heavily to position itself as a neutral AI hub, courting Washington with promised US investment while keeping trade and infrastructure links to China intact. Washington’s dilemma is whether that hedging strategy amounts to leakage risk or simply the price of competing with Gulf capital for compute build-out; the ongoing compliance monitoring built into the approval suggests officials see it as some of both, rather than a question already settled.

The chip curtain was drawn around a rival superpower; it is now being asked to police an ally hedging between two of them.

What happens next matters beyond this one approval. Every Gulf state, and a growing list of middle powers courting AI investment, is watching whether strategic ambiguity towards China disqualifies a buyer or merely earns closer supervision. With Saudi Arabia and the UAE both now cleared, subject to compliance, other governments have a template to cite in their own negotiations, effectively pricing Chinese-adjacency as a monitored variable rather than a hard barrier. Washington’s export control map, drawn as a line between friend and rival, is quietly turning into something closer to a gradient.

Imperial Valley Computer Manufacturing spent much of last year telling residents of California’s Imperial County that its planned data centre campus would leave the Colorado River alone entirely. The pledge appeared in a February blog post on the website of Sebastian Rucci, the Huntington Beach lawyer and developer behind the project, which said the facility would rely solely on recycled municipal wastewater and “does not touch a single drop of the Colorado River.” That pledge has now collapsed into a lawsuit.

After months of negotiations with the nearby cities of Imperial and El Centro over recycled-water supply broke down, and the Imperial Irrigation District rejected the company’s water application in early May, IVCM filed suit in Imperial County Superior Court on 5 June seeking roughly 260 million gallons of river water annually — enough, on the company’s own figures, to cover a few thousand county residents’ yearly use. Rather than simply asking for a fresh allocation, the suit pursues the water through a farmland-fallowing arrangement, paying growers to fallow acreage and transferring their rights to the project, a mechanism water-rights lawyers are watching as a possible template elsewhere. The facility would be California’s largest data centre, a 950,000 square-foot, 330-megawatt campus with its own battery storage, in a region already governed by some of the most contested water rights in the American West.

What makes the episode instructive is not the volume at stake, small against the valley’s vast agricultural allocation, but the sequencing, and what has followed. The commitment not to draw on the river predated the recycled-water deals it depended on; IVCM’s own suit calls the switch to the irrigation district a “last resort” after those deals collapsed. Days after a board candidate IVCM had backed lost the June primary for a seat on the district’s board, the company sued that same board, and county supervisors who fast-tracked the project in April have since reversed course, imposing a moratorium while a 19-member committee rewrites the rules by 2027.

The facility purchases municipal wastewater that would otherwise be discarded. It does not touch a single drop of the Colorado River.

That line, from the February blog post, now reads as a contingent promise rather than a design constraint. Whether the fallowing mechanism satisfies the valley’s water-rights holders, or hands the next developer a route to the same river, is what the 2027 rules will have to decide.

Vladimir Putin has moved to establish a national task force charged with coordinating Russia’s generative AI efforts, casting continued dependence on foreign large language models as incompatible with the country’s sovereignty. He has since gone further, signing a decree creating a presidential commission on AI development, converting the announcement into a standing structure with authority over funding and policy across ministries. The instinct itself is not new: sanctions and export controls have already cut Russian developers off from the latest Western chips and, increasingly, from the cloud infrastructure that trains frontier models. What is new is the explicit political framing: AI dependency is being treated less as a commercial inconvenience than as a strategic vulnerability on a par with energy or defence supply chains.

The practical starting point is thinner than the rhetoric suggests. Russian firms such as Sberbank and Yandex already run their own large language models, built with a mix of domestic engineering and hardware improvised around sanctions, but neither has demonstrated the scale or reliability of the systems now shipping from Western and Chinese labs. A coordinating commission can align funding, data access and procurement across ministries and state-linked companies, but it cannot conjure the advanced compute that remains the binding constraint on any serious model-training effort.

Declaring AI sovereignty a national priority is easier than building the compute base to make it real.

The more interesting comparison is with China, which has spent years cultivating open-weight models as a soft-power asset and has built a far more developed domestic chip supply chain than Russia’s, even though its own access to the most advanced chips is still constrained by Western export controls. Moscow’s version of the sovereignty argument arrives with none of that industrial base, which makes the initiative read as much as a signal of intent and a hedge against deeper technological isolation as a credible plan to close the gap with the AI frontier.

Amazon sued Perplexity AI in November 2025 over Comet, its agentic browser that can shop on a user’s behalf, alleging Comet accessed customers’ private Amazon accounts without identifying itself, ignored repeated demands to stop, and disguised its traffic as ordinary human browsing. A federal judge granted Amazon a preliminary injunction in March 2026; the Ninth Circuit has since stayed that injunction pending Perplexity’s appeal. The live question is no longer just whether Amazon can sue — it’s whether a customer’s authorisation lets an agent keep operating inside a platform’s accounts once the platform has expressly withdrawn permission.

Amazon frames this as a control problem as much as a disclosure one: customer consent, it argues, doesn’t automatically license outside software to operate inside its systems. Perplexity’s answer is that Comet is simply the customer’s chosen proxy, acting on authority the user already has. The ACLU has separately warned that stretching computer-fraud law to reach permitted tools like this one could cut far wider than a single shopping agent.

The lawsuit is less about who gets to click “buy” and more about whether a website can still tell who, or what, is doing the clicking.

The web has lived with unloved automation before — scrapers, price bots, browser extensions, many of them just as evasive as anything alleged here. What’s new is an agent operating inside an authenticated account, built to be indistinguishable from the person who owns it. Litigation and technical defences aren’t alternatives; both are already under way. The harder question is whether agents get a standard way to identify themselves, declare who authorised them and what they’re allowed to do — a web equivalent of a scoped access token — rather than every platform relitigating consent case by case.

Amazon’s decision to eliminate around 14,000 corporate positions was framed by leadership as a routine trim of management layers, the kind of bureaucracy-flattening exercise large companies periodically undertake. But the timing and the language around it, arriving alongside heavy internal promotion of generative AI tools for coding, customer service and logistics planning, made clear that this round was different in kind, not just scale.

What makes the Amazon cuts notable is not the number itself, which is modest against a workforce of well over a million, but the precedent. Executives openly linked the reductions to efficiency gains from AI adoption, making Amazon one of the first companies of its size to say so plainly rather than hide behind generic language about restructuring. That candor matters because it removes the plausible deniability other firms have relied on when announcing layoffs during an AI boom.

The test of an AI productivity claim is not the software demo. It is the org chart a year later.

The harder question is whether the efficiency is real or whether the layoffs are simply cover, a chance to trim costs while pointing at a convenient and fashionable explanation. Either way, Amazon has set a template other large employers will likely follow: cite AI, cut middle management, and let the ambiguity between genuine automation and ordinary cost discipline do the rest of the work.

What happens next is the actual test. If Amazon’s remaining teams sustain output with fewer people, the claim holds and other employers will move faster to copy it. If service quality or delivery timelines slip, the AI justification will look like what skeptics already suspect: a rebranding of an ordinary layoff for a market that rewards AI narratives.

On November 18, Cloudflare knocked out large parts of the internet for nearly six hours, not through an attack but through an internal mistake. A database permissions change at 11:05 UTC caused a query to return duplicate rows, doubling the size of Cloudflare’s Bot Management feature file. The malformed file then cascaded into failures across services that had nothing to do with bot detection, producing widespread errors on sites and apps that most users would not have known ran through Cloudflare at all.

That is the real story here, more than the outage itself. On November 18, 2025, at 11:20 UTC, Cloudflare—a critical content delivery network (CDN) and internet security provider serving 19.3% of all websites—experienced catastrophic service failures that cascaded across the internet for nearly six hours. Cloudflare sits between users and the origin servers of companies that never chose Cloudflare directly but inherited it through whichever vendor or platform they built on. When that layer fails, it does not fail as “Cloudflare is down.” It fails as X, ChatGPT, and Spotify going dark simultaneously, for reasons none of those companies control and few of their users could name.

A handful of infrastructure providers have become load-bearing walls for the entire web, invisible until they crack.

Cloudflare’s own account makes clear this was not sabotage. Cloudflare’s post-mortem makes it clear that the cause was an internal error – a latent bug triggered by a routine update – rather than an external attack. That is arguably more unsettling than an attack would be: it means the fragility is structural, waiting on the next routine deployment, at any one of the small number of firms the internet has quietly decided to depend on.

OpenAI closed out a year of legal wrangling by completing a corporate restructuring that its own board describes as settled and durable. OpenAI has completed its recapitalization, simplifying its corporate structure, with the nonprofit remaining in control of the for-profit and gaining a direct path to major resources before AGI arrives. The nonprofit arm, now called the OpenAI Foundation, holds a $130 billion stake in the for-profit enterprise, a public benefit corporation called OpenAI Group PBC. By any measure, that stake makes the Foundation one of the wealthiest philanthropic entities in existence, and the nonprofit’s equity in the for-profit is valued at approximately $130 billion, making it one of the best resourced philanthropic organizations ever.

The headline claim is that mission still outranks money. But the mechanism meant to enforce that claim looks thinner than the press release suggests. The for-profit’s board will be appointed solely by the overarching nonprofit entity and will, for now, consist entirely of the nonprofit organization’s board members. That is not independent oversight so much as the same people checking their own work under two different letterheads. A watchdog report tracking the restructuring put the concern more bluntly, arguing that OpenAI’s plan would leave the nonprofit continuing to transition to a new structure “without a duty to put its mission above the interests of shareholders.”

“I see that as a bit of an empty, unenforceable promise.”

That line, from a law professor commenting on the public benefit corporation statute governing OpenAI’s new structure, captures the gap between legal form and practical constraint. Referring to the laws surrounding public benefit corporations, the professor said the statute gives OpenAI wide latitude in deciding when to follow profit and when to follow its nonprofit mission. California’s attorney general extracted some concessions before standing aside, saying regulators “secured concessions that ensure charitable assets are used for their intended purpose, safety will be prioritized, as well as a commitment that OpenAI will remain right here in California.” Those are real guardrails, but they are external ones, imposed by a state regulator rather than built into the company’s own governance.

What OpenAI has actually built is a structure that satisfies the letter of “nonprofit control” while concentrating enormous financial upside and product authority in the same small group of people. The arrangement may hold for as long as that group agrees with itself. The test will come the first time the for-profit’s commercial interests and the Foundation’s stated mission genuinely diverge, and there is no independent board left to referee the disagreement.

China spent the past two years turning open-weight AI models into a foreign-policy asset, and it now appears ready to lock the door behind them.

According to reporting picked up widely this month, China’s Ministry of Commerce has spent weeks consulting Alibaba, ByteDance and Zhipu (rebranded Z.ai) on export controls that would restrict foreign access to the country’s most advanced model weights, training data and even chip designs. The mechanics matter: overseas users would keep access to these systems as cloud services, but downloading and running the weights themselves — the thing that made DeepSeek a global sensation and let developers everywhere fine-tune and deploy Chinese models on their own infrastructure — would require a license.

The line being drawn is not about who gets to use the models, but who gets to hold them.

That distinction is the whole story. Open weights were never just a technical choice; they were Beijing’s answer to a US-led chip embargo, a way to buy soft power and global developer mindshare cheaply while hardware remained scarce. Every enterprise building on a downloaded Chinese model, every research lab fine-tuning one offline, became a small vector for Chinese AI’s international standing, at essentially no cost to the state beyond the compute already spent on training. Restricting that flow reframes model weights the way Washington has long treated advanced lithography and chip IP: as strategic assets to be metered, not gifts to be distributed.

The irony is not subtle. Beijing spent years condemning US export controls as protectionist overreach designed to contain China’s rise. It is now proposing to apply the identical logic to its own frontier technology, presumably for the same reason Washington did: because openness, once a company or country is actually ahead, starts to look less like generosity and more like giving away the lead.

The quantum computing field has spent the past two years announcing one milestone after another, from error-corrected chips to topological qubits. Against that backdrop, a group of physicists recently set out to independently verify some of the discipline’s most exciting claims. What they found was not confirmation but erosion: careful replication showed that signals once celebrated as breakthroughs could be explained by simpler, more mundane physics.

The episode is notable less for the specific results it undercuts than for what happened next. The team’s findings, despite their significance, initially struggled to find a publisher, a reminder that the incentive structure in a hyped field rewards splashy claims far more readily than the quieter work of checking them.

Instead of confirming breakthroughs, the replication studies revealed that signals once hailed as major advances could actually be explained in simpler ways.

That tension matters beyond academic bookkeeping. Around the same time, papers from Google and a quantum startup argued that machines capable of breaking internet encryption could arrive sooner than expected, prompting Google to move up its own migration timeline to 2029, six years ahead of the federal deadline. Outside researchers were quick to note that the paper had not been peer-reviewed and rested on assumptions about future hardware that remain untested. Policymakers and enterprises are now being asked to plan cryptographic migrations against a scientific record where the boundary between genuine advance and inflated claim is still being sorted out, often after the fact.

African teams did really well, and had a lot of players born in Europe.

The immediate trigger is straightforward: A significant development in technology and policy.. But the longer pattern matters more. Where this lands in the next 12 months will shape whether other actors move the same way, or whether it becomes an outlier.

The first US operational use of one-way attack sea drones matters less as a one-off tactical event than as a signal of where naval competition is heading. For decades, maritime power has been measured in expensive platforms: carriers, destroyers, submarines, and the industrial base required to maintain them. One-way maritime drones introduce a parallel logic: cheaper hulls, faster iteration cycles, and missions that are politically easier to authorize because the system is expendable and no crew is at direct risk. That shifts part of the balance from prestige fleets toward production speed, software quality, and targeting integration.

The near-term military effect is not that drones replace major warships, but that they force every navy to spend more on defense against relatively inexpensive threats. A single successful strike can impose asymmetric costs if it damages critical sensors, logistics nodes, or traffic chokepoints. In practical terms, this pushes commanders toward layered defenses, persistent surveillance, and electronic warfare at sea, all of which increase operational complexity. It also narrows the margin for error in congested waters where commercial shipping, coast guards, and naval units already operate in close proximity.

There is also a strategic signaling effect. Once a major power normalizes one-way maritime drone use, peers and regional rivals are likely to treat it as doctrine, not exception. That can accelerate copycat procurement and local adaptation, including hybrid designs built from commercial components and custom guidance stacks. The result is a wider diffusion curve than traditional naval weapons programs, which tend to be slower and concentrated among richer states. In that environment, deterrence gets harder because attribution, intent, and escalation thresholds become less clear, especially when unmanned systems can be launched from deniable positions.

The long-term implication is that maritime security may start to resemble cyber competition: constant low-level probing, frequent adaptation, and recurring pressure on defense budgets to keep up with agile threat cycles. The US move may prove to be the opening step in that transition. If so, the decisive advantage will belong less to who owns the largest fleet, and more to who can update doctrine, software, and countermeasures fastest without triggering uncontrolled escalation.

Three years after Sundar Pichai reportedly called a “Code Red” inside Google over the threat ChatGPT posed to Search, Sam Altman has invoked the same language against Google. In a memo to OpenAI staff on December 1, Altman said he was declaring a “Code Red” to marshal resources toward improving ChatGPT as competitive pressure from Google and other rivals intensifies, and the message was reported to have urged employees to prioritize the advancement of its core model, delaying other initiatives to do so.

The immediate trigger was Gemini 3, which landed with enough force that even OpenAI-aligned executives said so in public. Salesforce’s Marc Benioff wrote that after using ChatGPT daily for three years, two hours with Gemini 3 changed his mind, calling the leap “insane” across reasoning, speed, images and video.

It feels like the world just changed, again.

What makes this notable is not the memo itself but the symmetry it exposes. The entire narrative of the last three years has rested on the idea that OpenAI possessed a durable technical lead, a moat built on data, talent and a head start that rivals would spend years closing. That framing is now inverted: the company that triggered the industry’s first panic is issuing its own, and doing so in language borrowed directly from the rival it once destabilized.

The deeper story is how thin any single company’s advantage in frontier AI has turned out to be. Model quality gaps that once looked structural are closing in weeks rather than years, and the leaders on any given month are increasingly reacting to each other’s releases rather than executing independent roadmaps. A “Code Red” is, by definition, an admission that the plan did not survive contact with a competitor’s ship date.

The Federal Energy Regulatory Commission has ruled that data centers can wire themselves directly to power plants rather than wait in line for a grid connection. On December 18, FERC found that PJM Interconnection’s existing tariff was legally deficient, ruling that it “does not include transmission services for an eligible customer to take service on behalf of a co-located load, and for that customer to be able to limit its energy withdrawals from the transmission system under certain conditions.” The order was issued after PJM’s own capacity auction closed short of its reliability target, a gap regulators have linked directly to surging AI power demand.

The practical effect is to legalize and formalize what several data center developers were already trying to do: sit an AI campus next to a nuclear or gas plant and draw power before it ever reaches the shared grid. FERC’s fix is procedural rather than symbolic. It orders PJM to create three new categories of transmission service, firm, nonfirm, and interim nonfirm, so that colocated customers can be billed and scheduled without simply riding free on services like black-start and frequency regulation that keep the wider system stable.

Co-location is framed by its regulators as a tool for meeting AI demand while minimizing the impact on the grid and consumers, not a workaround for the grid itself.

That framing is doing a lot of work. The order does not touch the harder jurisdictional question of who ultimately governs a load that never leaves the fence line of a power plant, an issue left for a separate rulemaking pushed by the Department of Energy. It applies only to PJM for now, but the commission has sent parallel directives to grid operators in New York, California, New England and the central United States, effectively asking every major market to explain why it isn’t doing the same thing. The queue that generators and ordinary consumers have waited years to clear is being redrawn, quietly, around whoever can get a data center built next to a reactor first.

On a July trade visit to Beijing, the Netherlands’ foreign trade minister declared the Nexperia affair effectively resolved, praising a restored dialogue with China after nine months of open acrimony. The governments have every reason to want it closed: the standoff, which began when The Hague invoked emergency powers in late 2025 to seize control of the Chinese-owned chipmaker over technology-transfer fears, triggered Chinese export restrictions that idled automotive production lines from Japan to Europe. What the ministers describe is real. What they are not describing is a company that has been put back together.

Because it hasn’t. Nexperia’s Chinese arm, backed by parent Wingtech, declared itself independent of the European headquarters months ago and now sources wafers from domestic suppliers rather than the group’s own network. Wingtech has sued for damages in the billions, and the two sides communicate, when they communicate at all, through statements accusing each other of disabling IT systems and building unauthorized parallel infrastructure. A single multinational has been cleaved into two firms that share a name and little else.

Diplomats can reopen a channel far faster than a corporation can reassemble a supply chain that has been deliberately torn in two.

That gap between the political settlement and the corporate reality is the story worth watching. Governments treat the crisis as a bilateral irritant to be smoothed over ahead of a wider EU-China trade fight; for the business, the separation is hardening into permanence, with each half securing its own inputs and its own customers. The lesson for other cross-border firms sitting on contested technology is unsentimental. When a company becomes an instrument of national security policy, it can be split along the same lines as the states fighting over it, and no amount of restored diplomatic warmth necessarily glues it back.

On the last day of 2025, Finnish authorities boarded and seized a cargo ship called the Fitburg, en route from Russia to Israel, on suspicion that it had dragged its anchor across an Elisa telecoms cable running under the Gulf of Finland to Estonia. Investigators later said the vessel was also carrying a hidden cargo of sanctioned Russian steel. It is the latest entry in a now-familiar Baltic sequence: a suspicious track near critical infrastructure, a damaged cable, a detained ship and a promise of accountability.

That promise is where the pattern keeps breaking down. In October 2025, a Finnish court threw out the case against the crew of the Eagle S, the Cook Islands-flagged tanker accused of severing a power cable on Christmas Day 2024, ruling that prosecutors had not proven intent and that any negligence claim belonged to the flag state or the crew’s home countries. The larger 2024 case, involving the Chinese bulk carrier Yi Peng 3, ended with the ship sailing on after a diplomatic standoff, its captain facing proceedings in Hong Kong rather than Europe.

The gray zone works precisely because it is deniable, and the law of the sea was not written for anchors used as weapons.

The Fitburg seizure suggests Baltic states are getting faster and more assertive at the interdiction stage, which is the part they can control from their own waters. What they still lack is a reliable way to turn a detained ship into a conviction, given flag-state jurisdiction, the difficulty of proving deliberate intent and the thin line between sabotage and clumsy seamanship. Until that gap closes, each seizure risks reinforcing the same lesson for anyone tempted to drag an anchor: the cost of being caught remains low, and the odds of being punished lower still.

The United States spent decades insisting that it did not pick winners; it merely set the rules of the game. That posture is harder to maintain now that the federal government has turned billions of dollars in Intel subsidies into roughly a ten percent equity stake, making the state one of the company’s largest shareholders. What was pitched as industrial policy has become, in effect, industrial ownership.

The logic is not hard to follow. Intel is the last American firm still trying to manufacture leading-edge logic on domestic soil, and its stumbles have national-security weight that a purely commercial calculus would never assign. Converting grants into shares lets Washington argue that taxpayers get an upside rather than a one-way handout. But it also blurs the line between regulator, customer and owner in an industry where the government already controls export licenses and buys the output.

Once the state holds the stock, every decision about a struggling champion becomes a political decision, whether or not anyone wants it to be.

The uncomfortable precedent is the point. A stake framed as passive rarely stays fully passive when a strategic asset falters, a plant is threatened, or a foreign buyer circles. Washington now has to answer a question it long avoided: not whether to support the chip industry, but how to behave as its co-owner without turning national champions into wards of the state.

The US Department of Commerce has told exporters that its licensing requirements for advanced AI chips now reach every business headquartered in China or owned by a Chinese parent, regardless of where that business is actually located. The move quietly changes the logic of the export regime: what matters is no longer the physical destination of a shipment but the corporate lineage of the buyer.

That shift closes an obvious workaround. For years the practical question was where a data center sat and whether a chip crossed a controlled border; a Chinese-owned subsidiary in a permissive jurisdiction could, in principle, buy freely. Treating ownership as the trigger collapses that gap, and it puts the burden on suppliers and their distributors to trace who ultimately controls a counterparty before a sale clears.

The border being policed is no longer a line on a map but a line on an org chart.

The cost of that clarity is complexity. Global supply chains are dense with joint ventures, holding companies and minority stakes, and a rule keyed to parentage forces exporters into due diligence that governments are better equipped to demand than firms are to perform. It also invites the predictable countermeasures — restructured ownership, opaque intermediaries, and further fragmentation of a chip market that Washington and Beijing are steadily pulling into separate orbits. The direction of travel is unmistakable: export control is becoming less about hardware and more about the identity of who stands behind it.

The European Union built its reputation on regulating technology first and asking the rest of the world to follow. The AI Act was the centerpiece of that strategy: a risk-tiered framework that banned a handful of uses outright, imposed transparency duties on general-purpose models from August, and reserved its most demanding obligations for high-risk systems due to take effect in 2026. Now, with a package of digital “simplification” proposals, the Commission is signaling it may push back and pare down precisely those high-risk provisions before they ever apply.

The official framing is competitiveness. European firms and their American suppliers have argued the compliance burden is heavy, the guidance is late, and the timelines assume a maturity the standards bodies have not delivered. There is truth in that. But a delay dressed as tidying-up is still a retreat, and it lands at the moment the law was supposed to prove it was more than a press release.

A rulebook that keeps postponing its own hardest chapters is not a standard; it is an intention.

The deeper cost is credibility. The “Brussels effect” only works if others believe the rules are real and durable enough to build around. Loosen the terms under industry pressure now, and every future negotiating partner learns that European deadlines are opening bids. The lesson for other governments drafting AI law is less about the specific clauses than about the sequence: writing the strictest regime in the world is the easy part, and enforcing it against your own economy is where the resolve actually gets tested.

In November, Anthropic disclosed what it framed as a watershed: a cyber-espionage operation, attributed to a Chinese state-sponsored group, that used its Claude models to automate the bulk of an intrusion campaign against roughly thirty targets across technology, finance and government. The company said the AI handled the overwhelming majority of the work — reconnaissance, exploitation, credential harvesting — with human operators stepping in at only a handful of decision points. It is the clearest account yet of an attacker treating a frontier model not as an assistant but as the labor.

The significance is real, and so is the caveat. Anthropic published the episode without the indicators of compromise or forensic detail that defenders normally expect, which left parts of the security community unable to independently verify the more dramatic autonomy claims. That gap matters, because the same report doubles as an argument for the company’s own products and for its preferred regulatory posture.

When the firm that sells the model is also the firm that discloses the breach, the incident report and the sales pitch begin to share a byline.

Strip away the framing and a durable point remains. The barrier to a competent intrusion has always been skilled human time, and tools that compress that time change the economics of who can attack whom and how often. Whether or not this particular campaign was as hands-off as advertised, the direction is not in dispute — and defenders should plan for the capability regardless of how loudly its vendor announces it.

On December 11, President Trump signed an executive order directing the federal government to break up the emerging patchwork of state artificial-intelligence regulation. It stands up a Justice Department task force to challenge state laws deemed burdensome, and it moves to condition billions in undistributed broadband funding on whether states fall into line.

The mechanism matters more than the rhetoric. Congress has not passed a federal AI statute, so the order reaches instead for the tools an administration already controls: litigation and the power of the purse. The Commerce Department is instructed to make states with laws judged “onerous” ineligible for remaining BEAD broadband money, quietly converting infrastructure dollars into leverage over legislation.

When the federal government cannot write the rules, it can still decide who pays for defying them.

The near-term target is not hidden. Colorado’s AI Act is named as the kind of law the task force might contest, and California’s newer measures sit within reach. With more than a thousand AI-related bills introduced across the states this year, the order is less a policy than a jurisdictional claim: that Washington should be the only venue where AI rules are written.

That claim is untested. Preemption without a supporting statute leans on contested readings of federal authority, and states retain broad power to legislate on consumer protection and civil rights. The likeliest outcome is not a settled national standard but a prolonged fight over who gets to set one, resolved less in legislatures than in court.

On 10 December, Australia became the first country to bar under-16s from holding social media accounts, placing the legal burden on the platforms rather than on parents. Services including TikTok, Instagram, Facebook, Snapchat, X and YouTube are required to take reasonable steps to detect and remove younger users, with penalties reaching tens of millions of dollars for those that do not.

The framing is child safety, and the political consensus behind it is unusually broad. But the mechanism is quieter and more consequential than the headline. You cannot reliably keep fifteen-year-olds off a platform without, in some form, establishing the age of everyone else — through document checks, facial-age estimation or behavioral inference. A rule aimed at a minority becomes an identity layer applied to the majority.

A law written to protect children ends up rewriting the terms on which adults use the internet.

Enforcement is where the intent meets friction. Age-assurance technology is improving but remains imperfect and straightforward to circumvent with a VPN or a borrowed login, which means the ban’s practical reach may trail its symbolic weight for some time. The more durable effect is the precedent. Regulators across Europe and beyond have floated comparable limits, and Australia has now demonstrated that a democracy can legislate a hard age line and dare the platforms to comply. Whatever the measurable results for teenagers, other governments will study the template — and the surveillance it quietly requires — long before the evidence on harm is in.

The most striking feature of the AI build-out is no longer the technology but the financing. Nvidia has said it intends to invest as much as $100 billion in OpenAI, whose spending in turn flows heavily toward Nvidia’s chips. OpenAI has reached a parallel arrangement with AMD, taking warrants for a sizeable stake tied to how many processors it agrees to buy, and has layered hundred-billion-dollar compute commitments on top with Oracle and others. Trace the flows and the same money can look like it is traveling a short loop among supplier, customer and backer.

This is not fraud, and vendor financing is not new; the telecom carriers of the dot-com era were propped up by the very equipment makers that sold to them, and the reckoning came when the demand failed to arrive on schedule. The concern now is structural rather than moral. When a chipmaker helps fund its largest customer to buy its own chips, demand and revenue begin to resemble the same figure counted twice.

A boom underwritten by a handful of interlocking balance sheets is only as durable as the weakest promise inside the loop.

The question worth watching is not whether the technology is real, but whether the committed compute is genuinely used and paid for out of end-customer revenue rather than recycled investment. Concentration is the quiet risk: a small cluster of firms is now both the market and its financiers, which means a stumble at one node propagates through the others. That is a fragility no amount of raw capacity can engineer away.

The story of the AI build-out has been told as a contest over accelerators: who ships the fastest processor, who books the most capacity. But Broadcom’s October launch of Thor Ultra, a networking part engineered to lash together enormous clusters of chips, is a reminder that the more durable advantage tends to sit in the connective tissue. Modern AI training is less a single supercomputer than a warehouse of processors pretending to be one, and that illusion is held together by the network.

Nvidia’s dominance has never rested on silicon alone. It rests on the proprietary fabric — the high-speed links and switching that let tens of thousands of chips move data as if they shared a single memory. Buy the accelerators and you are quietly buying the plumbing too, along with the switching costs that come with it.

The chip earns the headlines; the network decides who is locked in.

The counter-move is standardization. A coalition of Nvidia’s rivals is pushing open, Ethernet-based interconnects as the backbone for AI clusters, betting that customers spending tens of billions on infrastructure will eventually resent depending on one vendor for both the brains and the nervous system. Whether that bet pays off will not be settled by benchmarks. It will be settled by which architecture the hyperscalers standardize on over the next two years — a decision made in procurement meetings, far from the product launches, that will shape the economics of AI for the rest of the decade.

The economics of artificial intelligence were supposed to be a Silicon Valley story: capital expenditure, chip supply, model margins. Increasingly they are a utility story. In PJM, the grid operator that coordinates power for roughly 65 million people across the mid-Atlantic and Midwest, capacity prices in the latest auction climbed by around three-quarters, with data center demand cited as the primary driver. The cost of simply guaranteeing enough electricity to be available now runs into the tens of billions, and much of it flows through to ordinary ratepayers who never signed up to train a language model.

This is the part of the AI expansion that rarely makes the keynote slides. Hyperscalers negotiate favorable rates and long-term contracts; the residual cost of new transmission, standby generation and a tighter supply-demand balance tends to socialize across everyone else on the network. The result is a subtle transfer, from households and small businesses toward the balance sheets of a handful of the largest firms on earth.

The cloud does not float. It plugs into a grid that someone, somewhere, is paying to keep from failing.

The political friction is only beginning. States are starting to litigate who counts as a “large load” and what such customers should pay for the infrastructure their demand requires. Expect the fight over cost allocation, special tariffs for data centers, and the pace of new interconnections to become one of the defining domestic policy arguments of the AI era. The technology may be abstract, but the meter on the wall is not.