In partnership with
Last week, we covered Washington telling its diplomats to fight data sovereignty laws. This week, it started writing the enforcement mechanism.
Draft rules requiring US approval for every significant GPU export. A domestic AI company blacklisted for demanding weapons guardrails. Chip sales conditioned on diplomatic concessions and matching investments.
We've tracked the tightening of US control over AI infrastructure since the first China export rules.
This week, the pattern stopped looking like a series of one-off policy moves and started looking like a system.
I'm Ben Baldieri, and every week I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
The GPU Audio Companion Issue #95
Want the GPU breakdown without the reading? The Audio Companion does it for you, but only if you’re subscribed. If you can’t see it below, click here to fix that.
Today's issue is brought to you by packet·ai
The GPU runs on sponsorships. If you value independent, no-fluff analysis of the AI infrastructure market, the sponsors are who make it possible. I'd appreciate you checking them out.
Can efficiency and multi-tenancy really change GPU economics?
Packet·ai thinks so – and not just economics, simplicity for GPU buyers too.
With aggressive flat-rate pricing for enterprise-grade Blackwell, packet·ai targets AI startups and stealth projects that just want to build and launch, not fret over mounting hourly costs.
Find their new $199 flat-rate offer at packet.ai/blackwell.
It’s powered by the hosted·ai platform, profiled in The GPU #59.
US Drafts Rules to Control Every NVIDIA Chip Sale on Earth
Washington just wrote itself into every GPU procurement decision on the planet.
Bloomberg reports the Trump administration has drafted regulations requiring US government approval for AI chip exports to virtually every country. The framework creates a tiered licensing system: up to 1,000 GB300s, roughly a single-rack deployment, get simplified review. Larger clusters require pre-clearance with business model disclosure and US government site visits. Above 200,000 GB300s in one country, the host government negotiates directly with Washington and commits to "matching" investments in American AI. For context, 200,000 GB300s is the scale Nscale is planning for Microsoft across four US and European sites. The rules aren't finalised and could change or be shelved.
Why this matters:
Last week, we covered the State Department cable ordering diplomats to fight data sovereignty laws (Issue #93). This is the hardware enforcement mechanism. The cable targeted policy. These rules target silicon. Together they give Washington control over where AI infrastructure gets built and under what terms.
The tiered structure creates a de facto global compute hierarchy. France, India, Saudi Arabia, the UAE: every country planning gigawatt-scale AI facilities now factors in a diplomatic negotiation with Washington as a procurement dependency.
The administration that campaigned on deregulation is building the most interventionist AI governance framework any US government has proposed. Through executive authority, not legislation.
CoreWeave Signs Multi-Year Deal to Power Perplexity's Inference
CoreWeave is moving from training-heavy contracts into inference.
The neocloud has signed a multi-year partnership with Perplexity to run inference workloads on CoreWeave Cloud, powered by dedicated GB200 NVL72 clusters. Perplexity answers over 1.5 billion questions per month. CoreWeave's contract book to date has been training-dominated: the $9.7 billion Microsoft agreement, the OpenAI relationship. This deal marks a shift toward inference. Terms were not disclosed.
Why this matters:
Training contracts are large and capacity-reserved. Inference is latency-sensitive, usage-variable, and scales with end-user demand. 1.5 billion monthly queries is the kind of always-on production workload that tests whether a neocloud can deliver consistent performance, not just provision GPUs.
Perplexity is building tools for enterprise, and models replacing professional workflows generate inference compute at a scale that dwarfs chatbot queries.
The neoclouds that reliably serve production inference will capture a different margin profile than those that only rent training clusters. CoreWeave landing this contract, therefore, is strong validation of both their architectural choices, and an indication of the direction of travel for their future business model.
IREN Orders 50,000 B300 GPUs, Expands Fleet to 150,000
IREN is scaling faster than the Microsoft contract that put it on the map.
The NASDAQ-listed neocloud has entered purchase agreements for over 50,000 NVIDIA B300 GPUs, expanding its fleet to 150,000 across data centres in Mackenzie, British Columbia and Childress, Texas. The fleet is projected to support $3.7 billion in annualised revenue by year-end. Not fully contracted. IREN has secured $9.3 billion in funding over eight months across customer prepayments, convertible notes, GPU leasing, and GPU financing, with approximately $3.5 billion in additional capex expected in H2 2026.
Why this matters:
This is hardware to fill the $9.7 billion, 10-year Microsoft contract. 150,000 GPUs across two North American sites puts IREN alongside CoreWeave and Lambda in fleet size.
$9.3 billion through four non-equity capital sources. IREN is financing GPUs the way infrastructure companies finance physical assets: against contracted revenue and hardware residual value. Another $3.5 billion in H2 capex on top of that.
The $3.7 billion revenue target is not fully contracted. Hardware procurement moves faster than commercial agreements. The gap between fleet size and signed revenue is the risk every neocloud at this scale carries.
OpenAI Ships Its Most Capable Model Ever Into News Firestorm
GPT-5.4 might be the best model OpenAI has ever released. It launched into a news cycle that buried it completely.
On the same day OpenAI shipped GPT-5.4, scoring 83% on GDPval across 44 professional occupations, with 1M-token context, native computer-use, and 33% fewer hallucinations, the Pentagon designated Anthropic a supply chain risk, Dario Amodei called OpenAI's messaging "straight up lies," ChatGPT uninstalls surged 295%, and Claude shot to #2 on the App Store.
The sequence: Anthropic refused to let Claude be used for autonomous weapons or mass surveillance without explicit assurances. The Pentagon designated it a supply chain risk, a label normally reserved for foreign adversaries. Trump ordered a federal agency-wide ban. Anthropic had been the only AI provider operating inside the Pentagon's classified cloud. OpenAI stepped in with a contract allowing "all lawful purposes," claiming stronger guardrails through cloud-only deployment, a retained safety stack, and cleared personnel in the loop. Amodei called it "safety theatre." Anthropic is challenging the designation in court and has offered to continue providing models at nominal cost during the transition.
Why this matters:
GPT-5.4 is OpenAI's clearest move toward agents that replace professional workflows, not assist them. Matching professionals 83% of the time is a replacement metric. Agentic workloads at this capability level generate inference demand that dwarfs conversational AI. OpenAI's product team built something significant. OpenAI's deal team ensured nobody would talk about it.
OpenAI trusts the legal framework to hold. Anthropic doesn't, arguing that laws change and "lawful" is an expanding category under this administration. The consumer market, judging by uninstalls and App Store rankings, agrees with Anthropic.
The Pentagon designated an American AI company as a supply chain risk. Not for selling to adversaries, but for insisting on guardrails. Any enterprise running Claude in workflows adjacent to defence contracts must now evaluate whether the designation creates compliance exposure. Any AI company considering government sales now knows: accept all lawful uses without restriction, or don't bid.
Broadcom CEO Says AI Chip Sales Will Top $100 Billion in 2027
Hock Tan just put a number on the custom silicon threat to NVIDIA.
Broadcom's CEO told analysts the company has "line of sight" to $100 billion in AI chip revenue by 2027. Fivefold from $20 billion in 2025. AI revenue more than doubled to $8.4 billion in fiscal Q1. Quarterly revenue hit $19.3 billion; Q2 guidance came in at $22 billion versus $20.5 billion consensus. A $10 billion buyback was announced on top of $7.8 billion in Q1 repurchases. On the call, Tan named names: OpenAI shipping Broadcom chips at 1GW+ next year, Anthropic at 1GW this year and 3GW in 2027 via Google TPUs, Meta's custom accelerator roadmap scaling to "multiple gigawatts." Six new custom chip clients are expected this year.
Why this matters:
Broadcom is sizing its business in gigawatts, not units. Same language Meta used for its 6GW AMD deal last week. Competition now plays out at the facility scale.
OpenAI, Anthropic (via Google), Meta, and six unnamed clients are building Broadcom custom accelerators alongside NVIDIA GPUs. Meta's CFO confirmed custom chips capable of training, not just inference. Custom silicon is crossing from complement to competitor.
Tan's $100 billion target assumes his biggest customers keep scaling. One of them got blacklisted by the Pentagon this week. If Anthropic's designation disrupts its operations, it ripples through Google's TPU demand, which ripples through Broadcom's custom chip revenue.
Big Tech Warns Pentagon That Anthropic Designation Threatens Military AI Access
The tech industry just told the Pentagon it's overreaching. Without saying Anthropic's name.
ITI, the trade group representing NVIDIA, Amazon, and Apple, sent a letter to Defence Secretary Hegseth warning that the supply chain risk designation "creates uncertainty" and threatens military access to the best AI products. ITI's CEO argued that such designations are "typically reserved for foreign adversaries" and urged the Pentagon to work through the Federal Acquisition Security Council, the body created to evaluate procurement risk, rather than act unilaterally. OpenAI itself said Anthropic should not be designated.
Why this matters:
ITI's members include Anthropic's investors (Amazon), its compute suppliers (NVIDIA), and its competitors. The letter defends the principle that procurement disputes shouldn't trigger national security designations. It doesn't defend Anthropic's specific stance on weapons and surveillance. The worry is what this means for every other AI company negotiating with the government.
Lockheed and other contractors are reportedly already pulling Claude from their workflows.
Whoever fills the gap accepts the terms Anthropic refused. Six months to replace the only AI provider that operated inside the Pentagon's classified cloud.
US and Israel Strike Data Centres in Tehran. Iranian Drones Hit AWS.
Data centres are now military targets.
US and Israeli forces struck at least two data centres in Tehran. The strikes follow Iranian drone attacks on AWS facilities: two hit in the UAE, a third damaged in Bahrain, causing ongoing outages at ME-CENTRAL-1 and ME-SOUTH-1. AWS customers include the US Department of War. Whether military workloads were running alongside civilian is unknown. Iran enforced an internet shutdown, with connectivity at 1% of normal levels.
Why this matters:
The AWS strikes are the first publicly confirmed instance of a US hyperscale data centre being hit in a war. Every assumption that treated hyperscale facilities as civilian infrastructure immune from kinetic attack just got invalidated.
Every Gulf data centre deal we've covered sits inside this new reality. UAE chip approvals, Saudi Vision 2030, G42's partnerships: all assumed physical security as baseline. Drone strikes rewrite the risk premium on Middle Eastern infrastructure investment.
Insurance underwriters, site selection consultants, and hyperscaler risk teams are now pricing a variable that didn't exist in their models six months ago: whether the facility survives a regional conflict. That’s a big question for capital allocators, and could change the regional calculus in ways we cannot yet anticipate.
The Rundown
The US government had a busy week.
It drafted rules to approve every GPU sale on earth. It blacklisted a domestic AI company for setting terms on its own technology. It conditioned chip exports on diplomatic concessions. And it fought a war in which data centres, including American ones, became targets.
The market's response?
Keep building.
IREN ordered 50,000 GPUs. CoreWeave signed Perplexity. Broadcom projected $100 billion in custom silicon.
Capital is flowing toward infrastructure ownership and away from platform dependency, the same migration we tracked through sovereign AI in Issue #93, now accelerated by a week that made every assumption about jurisdiction, safety, and physical security harder to hold.
Whoever owns the physical layer owns the optionality. Washington is making sure it owns the rules of ownership.
See you next week.