In partnership with
On June 16, DOJ told a Memphis court that xAI's 57 unpermitted gas turbines were a matter of national security.
The same week, US export controls kept Anthropic's two top models offline.
Two federal statutes. One set aside. One enforced.
A productive week for national security.
Less so for frontier AI labs and local communities.
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.
Today's issue is brought to you by Rafay
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.
The World's Leading Neoclouds Run on Rafay
Rafay helps neoclouds, sovereign AI clouds, and enterprises turn GPU infrastructure into self-service, governed AI cloud services, from Token Factory and inferencing to Kubernetes, SLURM, bare metal, and VMs.
Anthropic's Top Models Pulled While xAI's Unpermitted Turbines Get Federal Defence
Two federal actions, one defending xAI, one restricting Anthropic.
The Department of Justice intervened in the xAI lawsuit on June 16, arguing that shutting down 57 unpermitted gas turbines at the Memphis data centre would undermine national, economic, and energy security. The filing claims Grok supports mission-critical military operations. The following day, DOJ moved to dismiss the Clean Air Act citizen suit entirely, defending the federal agencies named as co-defendants. On the export-control side, US restrictions remain in force against Anthropic's Fable 5 and Mythos 5, barring foreign nationals from access. Anthropic took both offline indefinitely. Calls for sovereign AI infrastructure outside US control followed within 48 hours.
Why this matters:
The structural read is selective enforcement. xAI's turbines violate Clean Air Act permits and are protected. Anthropic's models satisfy commercial law and are restricted. The criterion is alignment, not statute.
Anthropic's IPO timeline is the casualty. Fable 5 and Mythos 5 are the revenue layer that justifies the $65 billion Series H (Issue #108). Pulling them indefinitely punctures the pre-IPO narrative.
"AI sovereignty" is now a policy term, not a slogan. The UK Sovereign AI Fund and Era4-Cosine Lumen Sovereign (Issue #110) read as hedges against exactly this scenario. Capital wanting frontier-AI exposure without US export risk has fewer routes through US labs than it did a week ago.
SpaceX Buys Cursor for $60 Billion
The largest acquisition of a developer-tooling company on record landed a week after the largest IPO in history.
SpaceX acquired AI coding startup Cursor for $60 billion on June 16, company-stated and not independently verified. The deal closed the same week SpaceX shares traded as a public company for the first time, following the $75 billion IPO (Issue #110). Cursor's inference workload has historically run on Anthropic's Claude and OpenAI's GPT models. Whether it migrates to xAI on Colossus is the open question.
Why this matters:
Musk now owns compute (Colossus), models (Grok via xAI), foundry (Terafab with Intel and ASML, Issue #110), and developer tools (Cursor). No NVIDIA, no Anthropic, no OpenAI required at any layer. The vertical integration was capitalised in weeks.
Cursor was one of Anthropic's largest paying inference customers outside the hyperscalers. If it migrates onto Colossus, Anthropic loses revenue at the moment its top models are suspended (Top Story).
The price assumes consolidation, not standalone economics. Cursor's standalone valuation in earlier rounds was an order of magnitude lower. Every coding-assistant company is now a potential acquisition target for whoever owns the corresponding compute.
IREN Buys Nostrum Group for 490MW of Secured Spanish Power
490MW of secured, grid-connected Spanish power changed hands in one transaction.
IREN completed the acquisition of Spanish AI data centre developer Nostrum Group (Ingenostrum S.L.) on June 15, marking entry into the European market. The deal adds approximately 490MW of secured, grid-connected power in Spain plus an undisclosed development pipeline and a 50-person team. Nostrum founder Gabriel Nebreda joins. Operations continue under the IREN brand. Co-CEO Daniel Roberts cited Spain's renewables base and fibre connectivity. Terms undisclosed.
Why this matters:
Three months ago IREN issued the first A-rated securitised GPU debt at 3.31% blended (Issue #109). Today those proceeds bought Nostrum. CoreWeave, Applied Digital, and Crusoe are watching the same template clear.
Interconnection queues across continental Europe run five to eight years for facilities above 100MW. Nostrum's 490MW is already through the queue. IREN bought queue position, not greenfield.
Europe is writing sovereign AI cheques quarter by quarter: SoftBank €75 billion in France (Issue #109), UK £1.1 billion AI Hardware Plan and AMD £2 billion research commitment (Issue #110). Nebius holds UK (Issue #110). IREN now anchors Spain. The European footprint gets claimed in 2026, not later.
A note from Ben
Deliverance AI launched at London Tech Week on 8 June with a Sovereign Agentic Control Plane / OS. Zero to £6m ARR in three months. Probably the fastest-growing seed-stage startup in the UK.
They're hosting their launch event, Scaling the Sovereign Agentic Enterprise: From Governed Trust to P&L Impact at The 12th Knot, top floor of Sea Containers, on the 30th of June.
They've reserved a limited allocation for The GPU community. Apply on Luma and mention you came through The GPU.
China's AI Ecosystem Hardens Across Models, Capital, and Silicon
Three independent moves from China between June 16 and 17, hitting three distinct layers of the stack.
Z.ai released GLM-5.2 on June 16, a 753-billion-parameter open-weights model under MIT licence. It beats OpenAI's GPT-5.5 on long-horizon coding benchmarks at $5.80 per million tokens versus $35.00 for GPT-5.5, a 6x cost ratio in favour of the open-weights model. The same day, DeepSeek raised $7.4 billion, becoming China's most valuable AI startup. The following day, SCMP reported five AI models trained successfully on Chinese silicon alternatives to NVIDIA GPUs. The US Commerce Department refrained from adding DeepSeek to the Entity List the same week, while flagging over 100 other Chinese firms as security risks.
Why this matters:
Models, capital, and silicon all moved in 48 hours. Z.ai matches frontier US performance on coding at 1/6 the cost. DeepSeek captures the capital. Domestic silicon trains real models. The bottleneck thesis holds less weight every quarter.
The MIT licence on GLM-5.2 is the structural risk to US frontier labs. Open weights let enterprises self-host inference locally with no contract required. The lock-in that justifies frontier lab valuations is dissolving from below at the same time it is being broken from above by US export controls (Top Story).
US restrictions hit Anthropic and constrain Cursor's inference (Market Moves). DeepSeek raises $7.4 billion freely. Western labs face IPO risk, Chinese labs raise sovereign money. The market is open. The trade routes are not.
Tensordyne Tapes Out Napier on Broadcom Silicon
A 3nm inference chip with Broadcom as ASIC partner, claiming to beat Blackwell on power efficiency.
Tensordyne announced tape-out of Napier (TDN) on June 15, manufactured at TSMC 3nm with Broadcom as ASIC partner and HPE Juniper Networks on interconnect. The architecture replaces large-scale multiplication with addition-based logarithmic math, integrates SRAM tightly alongside HBM, and uses a proprietary any-to-any scale-up fabric. Tensordyne claims 17x tokens per watt and 13x throughput versus NVIDIA Blackwell. The company says one TDN rack (four 72-chip pods) matches nine NVIDIA Rubin plus Groq LPX racks on the same workload. Named production customers include Cirrascale and BlueSky Compute. LOI pipeline cited at $200 million. Series D anticipated later this year.
Why this matters:
Broadcom's ASIC business now serves as merchant silicon partner from the $35 billion Anthropic XPU build (Issue #110) down to pre-Series-D startups. Broadcom monetises every non-NVIDIA bet at every scale.
17x tokens per watt and 13x throughput are pre-silicon numbers, benchmarked against Blackwell rather than Rubin. The comparison rack mixes nine Rubin systems alongside Groq LPX. One named production customer. The Series D will price against the claims regardless.
Tensordyne's logarithmic math joins Fractile's analog compute (Issue #105) and Oriole's photonic interconnect (Issue #110), all positioned for the £150 million UK advance commitment for novel inference chips (Issue #110). The merchant inference market is becoming one.
AWS Builds the Agentic Stack on Its Own Silicon
AWS shipped agentic infrastructure at the top of the stack and ARM-shifted analytics at the bottom in the same week.
AWS announced four products at its New York Summit on June 17: Bedrock AgentCore (with Managed Knowledge Base) reached general availability, AWS Context launched as a knowledge graph service competing with Snowflake Horizon and Microsoft Fabric, WAF Bot Control launched as a managed bot mitigation layer, and AWS Continuum launched as a security service. The day before, AWS released Graviton-based Redshift RG instances, claiming 2.2x faster data warehouse and 2.4x faster data lake workloads at 30% lower price per vCPU than Intel-based RA3. Graviton5 (Issue #110) for agentic AI workloads followed two days later.
Why this matters:
AWS is building the agentic stack on its own silicon. Agents run on Graviton, talk to AWS Context, orchestrated by AgentCore. Each layer reduces dependency on NVIDIA. The non-NVIDIA hyperscaler thesis from Apollo and Broadcom (Issue #110) keeps compounding.
Redshift RG shifts analytics workloads off Intel x86 RA3 onto Graviton ARM at 30% lower cost. ARM is sufficient for the long tail of enterprise data warehouse work. GPU stays reserved for training and inference. CPU compute migrates from x86 to ARM beneath it.
AWS Context is the new enterprise lock-in vector. Three hyperscalers now compete for the same workload. Customers pick one, then default to that stack's compute downstream.
FERC Orders Grid Operators to Defend Large-Load Tariffs Within 30 Days
The biggest federal grid policy intervention for AI compute since the Three Mile Island restart.
The Federal Energy Regulatory Commission issued an order on June 18 directing regional transmission organisations to defend or revise their large-load interconnection tariffs within 30 days. The order specifically addresses AI data centre interconnections, where current tariffs and queue times have become the gating constraint on cluster deployment. NVIDIA's policy blog framed the order as removing the bottleneck for AI factories, fabs, and advanced manufacturing. PJM curtailment (Issue #107), Minneapolis grid pushback (Issue #108), and Crusoe pausing Project Jade in Wyoming "at the request of our customer" (Issue #110) all relate to the same constraint.
Why this matters:
The 30-day clock applies to PJM, MISO, ERCOT, and CAISO. The largest US data centre deployments sit in queues at these operators. Forcing tariff revision compresses interconnection timelines from five-plus years to whatever the revised tariffs allow. Capacity unbuildable in 2027 may become buildable in 2026.
The order is structurally similar to xAI's national security defence (Top Story). The federal government has decided AI infrastructure is critical infrastructure and is removing local frictions. State environmental challenges, grid queues, and permit fights are now subordinate to federal AI priorities.
Capital that moved into European secured supply this week (Neocloud) read the US market correctly. Even with federal intervention, US grid interconnections will take time to clear. European secured capacity remains the faster path to GPU deployment for the next 18 months.
The Rundown
The same federal apparatus gave opposite answers. Alignment decided both.
China built into the gap. Three layers moved at once. Models, capital, silicon.
Capital responded. European secured power, vertical integration into developer tools, agentic stack off NVIDIA. None of these bets assumed a global market.
FERC ordered the US queue cleared in 30 days. State frictions are now subordinate to federal AI priorities. Same logic as the xAI defence. Scaled.
NVIDIA still ships the most chips. The customers no longer share an ecosystem.
Another productive week for national security.
See you next week.