AI has become rule-change-driven. Innovation still matters, and so does surviving the next update to what counts as \u201callowed\u201d \u2014 on chips, on deals, on data, on deployment. In 2026, the most valuable thing in tech is no longer a model or a GPU. It\u2019s permission \u2014 the kind you can lose between a purchase order and a shipping label.<\/p>\n
That permission now shows up in places executives used to treat as boring. Nvidia is requiring Chinese buyers of its H200 AI chips to pay in full up front, with no cancellations, refunds, or last-minute configuration tweaks, according to Reuters<\/a>. In a few cases, collateral or commercial insurance can substitute for cash. It\u2019s a payment term that doubles as an admission: Policy volatility has become a line item, and the invoice is due before the truck leaves the dock.<\/p>\n Then came the other kind of paperwork. China\u2019s Ministry of Commerce said it would assess and investigate Meta\u2019s acquisition of AI agent company Manus<\/a>, pointing to compliance obligations around foreign investment, technology exports, and data transfers abroad. Manus is Singapore-based but has roots in China, and has said it has more than $100 million in annual recurring revenue \u2014 big enough to matter, small enough to be a useful test case.\u00a0<\/p>\n The business problem underneath all of this keeps getting messier, showing up in the connective tissues that executives can\u2019t model. The answer to \u201cCan you\u2026?\u201d keeps changing. Can you sell U.S. chips \u2014 then ship them? Can you buy a China-linked company \u2014 and keep it? Can you train, host, and deploy across borders without waking up to a new definition of what \u201ccounts\u201d as tech transfer? AI companies still compete on speed. They now also compete on durability \u2014 how well their plans survive the rewrites.<\/p>\n Chips have become paperwork<\/p>\n If chips were still a normal product category, Nvidia wouldn\u2019t need to behave like it\u2019s selling something between sanctioned equipment and BTS concert tickets. Yet the H200 trade now comes with ration math and political weather.<\/p>\n The Nvidia chips issue between the U.S. and China is supposedly about silicon. But it keeps turning into a permissions system with inventory attached. Nvidia CEO Jensen Huang said this week that Chinese customer demand for H200s is \u201chigh \u2014 quite high\u201d and that the company has \u201cfired up our supply chain\u201d accordingly. But now, Beijing is reportedly asking Chinese tech companies to stop shopping. Chinese firms have reportedly ordered more than 1.2 million H200 chips priced around $27,000 each, while Nvidia\u2019s inventory is closer to 700,000 units. But Chinese authorities have been directing companies to temporarily pause orders<\/a> while Beijing decides what ratio of domestic to imported chips buyers should maintain. That\u2019s demand, filtered through policy \u2014 a government shaping how much foreign advantage it wants inside its own data centers.<\/p>\n Nvidia\u2019s insistence on cash up front reads differently once you remember what a rule change can cost: The company took a $5.5 billion write-down in April<\/a> tied to a U.S. export ban. (And there have been more than a few export bans<\/a> for the company to contend with.) In a normal market, you can misread customers and recover. In this AI one, you can misread regulators and end up eating inventory that was \u201cstrategic\u201d only after you manufactured it.<\/p>\n China has been widening the perimeter below the headline battles, pushing the question down from \u201cwhich chip can you buy\u201d to \u201cwhich chip can you build your national infrastructure on.\u201d In November, Beijing issued guidance barring foreign-made AI chips from data center projects funded even partly by the state, forcing early-stage projects to remove or cancel plans to use them. At the time, Nvidia\u2019s share of China\u2019s AI chip market fell from 95% in 2022 to zero in 2025<\/a> \u2014 making it clear that this isn\u2019t a theoretical future.<\/p>\n \u201cIf the purchase orders come,\u201d Huang said this week at CES, \u201cit\u2019s because they\u2019re able to place purchase orders.\u201d Nvidia reportedly told Chinese clients that it aimed to begin shipping the H200 chips before the Lunar New Year holiday in mid-February, contingent on Beijing\u2019s approval.<\/p>\n The U.S. side keeps moving, too, and the movement itself has become part of the risk. In early December, President Donald Trump (controversially<\/a>) said the U.S. would allow exports of Nvidia\u2019s H200 chips to China<\/a> with a 25% fee on sales, under Commerce oversight. Then, the administration launched a review of advanced Nvidia chip sales to China after that announcement, showing how quickly the \u201callowed\u201d category can become conditional again. In practice, the corridor is never simply open or closed; it\u2019s perpetually subject to redefinition.<\/p>\n Deals have become tech transfers<\/p>\n When chips become strategic infrastructure, ownership starts looking strategic, too. China\u2019s Commerce Ministry framed its interest in the Meta\u2013Manus transaction around foreign investment, technology exports, and data moving abroad \u2014 the language of sovereignty, delivered through compliance. Manus being Singapore-based with roots in China is the key here, because it captures the modern corporate bet: that geography can be managed through incorporation, staffing, and paperwork. Beijing is signaling that those moves don\u2019t guarantee insulation.<\/p>\n Meta has said that there will be no continuing Chinese ownership interest after the deal and that Manus will discontinue services and operations in China. Even if that\u2019s enough to satisfy politics in both the U.S. and China, it doesn\u2019t automatically answer the question of what counts as an export when the \u201cthing\u201d being exported is a team, a system, and the operational know-how to build agents that can touch real workflows \u2014 strategic development, not normal software.<\/p>\n The broader implication is that AI acquisitions are increasingly treated less like ordinary M&A and more like capability transfer \u2014 the kind that governments want to inspect, slow, condition, or block. Europe is tightening the same screws in a more procedural register. EU lawmakers reached a provisional political agreement in December to reinforce the bloc\u2019s foreign direct investment screening regime. And EU regulators will decide by Feb. 10 whether to clear Alphabet\u2019s $32 billion acquisition of cybersecurity firm Wiz or open a deeper probe \u2014 a reminder that the \u201cAI stack\u201d includes the security layer that determines who trusts the cloud where models run.<\/p>\n And data has been put behind borders<\/p>\n The most durable rule changes don\u2019t arrive as a single ban. They arrive as definitions, frameworks, and timelines that force companies to build today around constraints that harden tomorrow. The European Commission\u2019s AI Act timeline is explicit: prohibited practices and AI literacy obligations began applying Feb. 2, 2025, obligations for general-purpose AI models apply from Aug. 2, 2025, and the Act becomes fully applicable Aug. 2, 2026, with some longer transitions extending to 2027. When companies asked Brussels for a delay, Reuters quoted a Commission spokesperson saying there would be no \u201cstop the clock.\u201d This \u2014 predictable, public, and still relentless \u2014 is the EU version of pressure.<\/p>\n Brussels is also aiming at the substrate AI rides on. In November, the European Commission opened market investigations under the Digital Markets Act into whether AWS and Microsoft Azure should be designated gatekeepers for cloud computing services. If cloud becomes a chokepoint under competition law, \u201cAI regulation\u201d stops being only about model behavior and starts looking like infrastructure governance \u2014 where deployment, logs, training data, and distribution power live.<\/p>\n