President Trump Calls AI Oversight a "Globalist Scheme" at the UN, Two Days Before Altman and Amodei Take Their Case to the Security Council; DeepSeek Bets Its Next Models on Huawei Chips as Opus 5.5 Cuts Price 40% to Meet Grok 4.7

On September 22, President Trump told the UN General Assembly that international AI oversight was a globalist scheme, said he would not let the growth of something bigger than the industrial revolution be held back, and lumped safety advocates in with climate-change alarmists. Less than 48 hours later, on the afternoon of September 23, OpenAI's Sam Altman and Anthropic's Dario Amodei each addressed the UN Security Council, calling for global AI safety standards and an incident-reporting mechanism; Amodei pledged that Anthropic would slow its release pace as needed — a unilateral commitment that binds no one but itself. The same week, President Xi Jinping made a state visit to Washington, and Nvidia's chip sales to China remained at zero: the H200 was approved last December, but no chips have shipped to China to date, because U.S. rules require the chips be used only within China while Beijing directs its own companies to use Nvidia chips only for overseas operations — two mutually exclusive rules, with neither side budging. On the Chinese side, DeepSeek founder Liang Wenfeng told investors that shifting next-generation model training to Huawei and domestic chips is now the company's top priority. The structural shift of the past seven days is this: the fight over AI governance no longer stays confined to lab statements and congressional wrangling — it has moved onto the podiums of President Trump himself and the United Nations, with both positions now fully on the table and neither side persuading the other. Meanwhile, the supply side of compute keeps moving along its own separate tracks — American chips stuck in political rules, Chinese chips stuck in capacity and efficiency — while the pricing war at the model layer hasn't waited for either of those to resolve.

A hoax at the General Assembly podium, and a plea at the Security Council

On September 22, President Trump addressed the UN General Assembly and called international AI safety oversight a globalist scheme, saying: "I'm not going to stifle the growth of something that will be bigger than the industrial revolution." He grouped those warning about AI risk together with climate-change alarmists, having already called existential-risk claims a hoax on Truth Social, and said the only safety measure needed was a sufficiently high-IQ president — namely himself. This tracked his stance on September 14, when he called Nvidia CEO Jensen Huang to say opposition to data center construction was a hoax and that critics were simply doing rivals' bidding.

Less than 48 hours later, on the afternoon of September 23, Altman and Amodei each addressed the UN Security Council, with Amodei appearing by video. Amodei's remarks worked on two levels: AI's potential benefits for science and medicine, and the risks of bioweapon development or systems slipping out of control. He pledged that Anthropic would slow the pace of new releases as needed to ensure safety — but this is a unilateral statement that binds no competitor or government, specifies no concrete risk threshold that would trigger a delay, and establishes no external body with veto power over release decisions; he also proposed global model-testing standards and a ban on bioweapon development, both advisory in nature. Altman, for his part, called for a "rapid incident reporting" mechanism and "global agreements on safeguards and oversight," stressing that any standard must not lock in incumbents or favor one business model, and must support open- and closed-weight developers alike, new entrants as well as established labs. Also speaking at the session were Hugging Face CEO Clément Delangue, who argued decisions shouldn't be driven by fear, and Microsoft President Brad Smith, who backed independent evaluators and human-control mechanisms.

This marks the first time the debate that began with Amodei's September 12 essay proposing a slower frontier pace has reached the UN podium — and the first time it has been directly rejected by President Trump himself, on the same international stage, within less than two days. For enterprises, the takeaway hasn't changed: who sets the pace remains unanswered; what can actually be acted on are each lab's voluntary disclosure and evaluation commitments, plus whatever audit clauses eventually make their way into contract addenda. And which models those clauses cover, and who verifies them, will depend on the terms enterprises negotiate with their own vendors — not on what got said at the UN.

President Xi Jinping's two days in Washington, Nvidia's China chip sales still at zero, DeepSeek bets its next models on Huawei

The same week, President Xi Jinping made a state visit to Washington, with OpenAI's Altman, Nvidia's Jensen Huang, and Apple chairman Tim Cook attending related state-dinner events. U.S. Trade Representative Jamieson Greer stated plainly that semiconductor controls "weren't even discussed" during the summit. In Congress, Senate Minority Leader Schumer demanded President Trump "draw a hard line," accusing him of handing AI technology to China; Senator Warren argued policy should be set through legislation rather than executive agreements, and criticized the deep involvement of executives like Huang in the negotiations. Three proposed export-control bills have been folded into the NDAA manager's amendment but remain stalled over disagreements on the Iran war and Pentagon spending; every major AI company and chip supplier except Anthropic opposes the bills, saying they would hamper innovation. The White House's response is that the current regime is already "the most stringent export-control system in modern history."

Nothing about the chip reality changed because of the summit. Nvidia's H200 received conditional U.S. export approval last December, with roughly ten Chinese companies — including Alibaba, Tencent, ByteDance, and JD.com — each licensed to import up to 75,000 units, and Lenovo and Foxconn designated as distributors, yet not a single chip has reached China: U.S. rules require the chips be used only within China, while Beijing directs its companies to confine Nvidia chip use to overseas operations and favor domestic chips instead — two mutually exclusive rules that neither side can satisfy. Nvidia's quarterly China revenue share has fallen from more than 20% to about 5%, and its share of China's AI-accelerator market, in Jensen Huang's own words, is now "close to zero."

It's against this backdrop that DeepSeek founder Liang Wenfeng told investors that shifting next-generation model training to Huawei and domestic chips is now the company's top priority. Huawei plans to deliver new batches of custom accelerators to DeepSeek between Q4 2026 and Q1 2027, with the Ascend 950DT expected before year-end and the Ascend 960DT debuting in Q1 2027 — consistent with the roadmap, three quarters ahead of schedule, that Huawei rotating chairman Eric Xu announced last week. DeepSeek is training a 2-trillion-parameter model, about 1.5 times the size of its released V4 (1.4 trillion parameters), with an eventual goal of 8 trillion parameters. Liang's efficiency comparison is blunt: training an OpenAI-class model takes roughly 50,000 Nvidia GB300 chips, versus about 200,000 Ascend 950 chips — four times as many. He believes Huawei's AI chips will catch up with Nvidia's within a few years, but for now this is a trade of scale for autonomy: four times the chip count in exchange for a training path free of export-control exposure.

The implication for enterprises is direct: compute supply on both sides of the U.S.-China divide is no longer governed by a single variable. American chips are hostage to two mutually exclusive usage rules with no near-term resolution; Chinese chips are hostage to absolute capacity and per-unit efficiency, with scaling up the only viable way to close the generational gap for now. Both paths point to the same conclusion: the cost structure of compute at the model layer won't stabilize over the next year or two just because of one summit or one bill.

Opus 5.5 cuts price 40%, Grok 4.7 expands at the same price, Gemini 4 forced to show its hand early

On September 21, xAI released Grok 4.7, keeping the same pricing as Grok 4.6 — $2 per million input tokens, $6 per million output tokens — while shipping an entirely new, larger base model rather than a fine-tune of 4.6, expanding the context window to 500K tokens and strengthening long-context self-verification. On Terminal-Bench 4.0, its score jumped from 4.6's 20.3% to 38.0%, and it led on the Harvey legal-agent benchmark and EEBench as well; in safety testing, only 3.3% of risky dual-use prompts got through. The very next day, September 22, Anthropic released Claude Opus 5.5, cutting input pricing to $4 per million tokens and output to $20, each down 20% from Opus 5, with cached-read pricing down 60% and a 40% reduction in total cost on typical workloads. Anthropic called it "the new leading model," citing 66.4% on Terminal-Bench 4.0 (versus 52.3% for Opus 5), a knowledge-work Elo score of 1846 (versus 1708), and an 85% drop in boundary-circumvention attempts.

Within two days, the two labs responded to the same pressure in opposite ways: xAI held price flat and gave more performance via a bigger model, while Anthropic held its positioning flat and cut price by 40%. This pricing war pushed Google straight into the spotlight. On September 23–24, at The Information's AI Agenda Live summit, Koray Kavukcuoglu — promoted to Senior Vice President of Google DeepMind in August — revealed that Gemini 4 has entered early post-training, with the goal of shipping "as soon as possible," well ahead of the year-end target set earlier. Gemini 4 began pre-training on July 21 and reached post-training within two months, faster than Gemini 3.5 Pro's prolonged testing cycle since June. But the gap remains wide: on the Intelligence Index, Gemini 3.6 Flash scores 34.0, GPT-6 Sol scores 47.5, GPT-6 Astra scores 52.7, and Opus 5.5 scores 57.6 — a 23-point gap between Google's current frontier and Anthropic's. The accelerated release looks less like an autonomous choice of pace and more like something forced by the September 21–22 pricing war.

Three product lines chose three different moves — price cut, capacity expansion, and an early sprint — in the same week, which tells you the pricing and performance rankings of frontier models are now shifting weekly rather than quarterly. Enterprises that lock model choice into a contract will quickly find themselves budgeting against a stale price list and stale benchmarks. AgentsFlare's multi-model routing dynamically dispatches by task type, cost budget, and SLA, so a reshuffled ranking at the model layer doesn't require re-evaluating contracts one by one — the routing policy updates as the price list does.

Oracle issues a force majeure notice on its New Mexico data center; the $18 billion loan trades at a 10% discount

On September 24, Oracle issued a force majeure notice on its Project Jupiter data center in New Mexico, citing continued delays from water- and air-permit litigation, a forced redesign of the power plan from gas turbines to Bloom Energy fuel cells, and financing-side pressure and uncertainty. The 1,400-acre campus was meant to support Oracle Cloud Infrastructure and OpenAI's Stargate compute needs, originally targeting deliveries starting in the first half of 2027; the timeline now faces slippage, and contractors, equipment suppliers, lenders, and OpenAI all face fresh uncertainty. The project broke ground in September 2025, and despite thousands of construction workers on site, was only about 26% complete as of early September 2026. New Mexico's Supreme Court rejected an emergency injunction request from environmental groups on September 17, but the broader permitting litigation remains unresolved.

The roughly $18 billion syndicated loan behind the project is now trading at 89 to 91 cents on the dollar in the secondary market, implying the market has priced in $1.6 to $2 billion of potential loss — partly a result of Oracle's credit rating being downgraded to BBB-, just one notch above junk, this past July. Oracle officially maintains the project remains on schedule; its fiscal 2026 cloud revenue reached $34 billion, up 39% year over year. This is the second signal in the past week reminding enterprise buyers that there's a meaningful gap between compute commitments on paper and actual delivery — and that gap itself is now being repriced by capital markets. For enterprises that have written compute guarantees into procurement contracts, execution risk tied to a single vendor or a single campus deserves the same budget buffer as a model-layer price cut cycle. AgentsFlare's cross-cloud routing lets the same policy span AWS, Azure, and private deployments, so a delivery delay at one region or one vendor doesn't drag down an enterprise's entire line of business.

Upstream compute: AMD's market cap tops $1 trillion, cloud compute prices rise 17%–41%

On September 21, AMD's market capitalization passed $1 trillion, making it the fourth U.S. chipmaker to cross that threshold, driven by factors including Meta's Muse assistant topping the free charts on Apple's App Store, strong memory-chip demand, and record South Korean chip exports. The same day, an unverified supply-chain channel report claimed AMD plans to raise prices roughly 10% across AI accelerators, consumer GPUs, and motherboard chipsets starting in Q4, attributing it to rising wafer costs at foundry partner TSMC; AMD has not confirmed the report, but its stock rose as much as 5% that day, with Nvidia and TSMC climbing in tandem and the Philadelphia Semiconductor Index up 2%. The price-hike rumor remains unconfirmed, yet the market already traded it as if it were real — a sign that tight chip supply is now the market's default assumption, one that moves stock prices without needing official confirmation.

That price transmission has already reached cloud bills. GPU cloud provider Nebius announced on September 17 that it will raise service pricing starting October 1: GPU prices up 17% to 21%, CPU up 25%, and memory up 41%. TrendForce's earlier monthly bulletin noted that HBM contract negotiations remain deadlocked because next-generation compute platform specs haven't been finalized, leaving clear room for prices to rise further, while its forecast for PC DRAM contract-price increases has also been revised upward. The price-hike expectations on the chip side and the price hikes already landed on the cloud side point to the same transmission chain: tight foundry costs and memory supply are moving from chip price sheets, through cloud-provider bills, into enterprises' monthly compute spend — and that chain hasn't yet reached model-layer API pricing. For enterprises that set token budgets by team and project, every upstream price increase means re-checking which team is using which pricing tier and where the actual cost is going; AgentsFlare's quota management and four-way cost attribution — by endpoint, user, team, and model — exists precisely so that check doesn't have to wait until the monthly bill arrives.

Special insight: inference cloud valuations doubled in three months

Baseten is a U.S. company that provides inference deployment and operations for companies without an internal infrastructure team, with customers including Abridge, Clay, Cursor, OpenEvidence, Mercor, and Notion. On September 22, Axios reported, citing sources, that Baseten is in talks with investors at a valuation of roughly $26 billion — just three months after its last round in June at about $13 billion; before that, in February, it raised a $300 million Series E at a $5 billion valuation, led by IVP and CapitalG with Nvidia participating. Modal Labs is a code-first platform for compute workloads that lets developers run code in the cloud without managing infrastructure themselves, with annual recurring revenue of about $50 million in early 2026. Around the same time, market reports indicated Modal is in talks for a new round at roughly a $15 billion valuation, up from its last publicly disclosed round — an $87 million Series B in September 2025 at a $1.1 billion post-money valuation — an increase of roughly 14x in under a year; a separate report of a roughly $5 billion valuation had surfaced about four months earlier. Both rounds remain in discussion, with investors and final terms not yet disclosed.

The two companies' businesses aren't the same: Baseten operates the managed-inference operations layer, while Modal operates a lower-level compute execution environment — but capital markets are sending the same signal: the layer that turns a model from code into a reliable service is now being priced as infrastructure on par with the model itself. On traction, Baseten already counts several well-known enterprise customers as users, and Modal has a confirmed annual revenue run rate; neither valuation rests purely on narrative. But doubling — or in Modal's case, rising more than tenfold — within three to four months also suggests pricing at this layer is currently driven mainly by private-market sentiment rather than revenue multiples, and it's worth watching whether the next round actually closes and whether the terms hold up to today's asking price.

Whose pace, whose chips, whose bill

Over the past week, three developments each exposed the real state of one layer of AI infrastructure. At the governance layer, the UN podium proved one thing: neither President Trump's hoax framing nor a lab's voluntary pledge can unilaterally set the pace of AI development on its own — what can actually be acted on are the audit clauses enterprises negotiate with their own vendors. At the compute layer, chip supply on both sides of the U.S.-China divide is stuck, respectively, in mutually exclusive political rules and a generational efficiency gap, and neither side can route around its own constraint in the short term; DeepSeek's choice to trade four times the chip count for autonomy is, at bottom, trading time for space. At the capital layer, model pricing and performance are being reshuffled weekly while inference-cloud valuations double quarterly — together showing that this industry's pace of asset repricing has outrun the budget cycles most enterprises run internally.

None of these three layers moves at a pace enterprises control. What enterprises can control is just one thing: putting the variables that matter to them — which model, how fast they can switch, where every dollar goes — into a system they themselves control, rather than locking them into any single vendor's roadmap.