Anthropic just rented the entire GPU cluster Elon Musk built to compete with it.
Then it doubled Claude Code's rate limits. And raised Opus API caps. And disclosed a compute stack totalling over 10 gigawatts across five providers.
Renting compute at that scale to a competitor is a new one.
But then I guess the enemy of my enemy is my friend?
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 #105
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.
Anthropic Rents Colossus from SpaceX
Colossus was built for Grok. Now, it’s running Claude.
Anthropic announced a partnership with SpaceX to use all compute capacity at Colossus 1 in Memphis: over 300 megawatts and 220,000 NVIDIA GPUs, coming online within the month. The deal doubled Claude Code rate limits, removed peak-hours throttling, and raised Opus API limits. Anthropic also disclosed interest in partnering with SpaceX to develop "multiple gigawatts of orbital AI compute capacity." The full committed stack now spans Amazon (5GW), Google/Broadcom (5GW of TPUs from 2027), Microsoft/NVIDIA ($30B Azure), Fluidstack ($50B US infrastructure), and SpaceX (300MW).
Why this matters:
xAI built Colossus as the world's largest single GPU cluster to train Grok. Grok trails every frontier coding model. Leasing the entire facility to Anthropic is the closest thing to a public concession that the cluster is worth more as rental income than as a competitive weapon.
Anthropic's $30 billion in annual revenue (Issue #103) created an inference capacity crisis that SpaceX's underutilised cluster could solve immediately. SpaceX gets revenue from hardware that failed to produce a competitive model. Anthropic gets 220,000 GPUs without breaking ground on anything.
No other AI lab rents from this many providers at this scale without owning any of it. OpenAI builds its own data centres. xAI built Colossus. Anthropic rents from everyone. The risk is dependency on partners who are also competitors (Amazon invested $33B, Google $40B). The advantage is speed: 220,000 GPUs online within the month, no construction required.
NVIDIA Takes $2.1 Billion in IREN
The vendor-customer relationship just became an investor-partner relationship.
NVIDIA and IREN announced a strategic partnership to deploy up to 5 gigawatts of DSX-aligned AI infrastructure. IREN issued NVIDIA a five-year right to purchase up to 30 million shares at $70 per share, a potential $2.1 billion equity investment. The flagship deployment will be IREN's 2-gigawatt Sweetwater campus in Texas.
Why this matters:
NVIDIA used to sell GPUs to neoclouds. Now it is taking equity across the entire AI infrastructure stack. SEC filings show $2 billion into Nebius (Issue #96), $2 billion into Synopsys (chip design tools), and $1 billion into Nokia (networking). Now $2.1 billion into IREN.
IREN's SEC filings show four data centres under construction in Childress, Texas (200MW, $2.4 billion in contracted Microsoft revenue) and $3.9 billion raised in convertible notes and equity in December 2025. NVIDIA is not betting on a speculative neocloud. It is backing a company with $2.4 billion in contracted Microsoft revenue and $3.9 billion in fresh capital.
The DSX flagship at Sweetwater is 2 gigawatts on a single campus. NVIDIA is co-designing the facility architecture, taking equity exposure, and using IREN as the reference deployment. Every neocloud building at gigawatt scale will be benchmarked against what comes out of Sweetwater.
DigitalOcean Launches AI Cloud, Beats AWS Bedrock by 3.9x
The third option between hyperscalers and bare-metal GPU rentals just shipped.
DigitalOcean launched its AI-Native Cloud at Deploy 2026: owned GPU infrastructure (NVIDIA H100, H200, B300, AMD MI300X, MI350X, MI355X across 20 data centres), an inference engine with 70+ models, managed databases with pgvector, and production agent runtimes. Character.ai runs 1 billion+ queries per day on the platform. Independent benchmarking by Artificial Analysis measured 3.9x higher output speed than AWS Bedrock on DeepSeek V3.2.
Why this matters:
DigitalOcean owns the silicon and co-engineers at the kernel level with both NVIDIA and AMD. That depth of partnership is one few neoclouds can match.
Character.ai at 1 billion queries per day is not a proof-of-concept. The 3.9x speed advantage over Bedrock was measured by an independent third party. DigitalOcean prices a representative agentic workload at $67,727/month versus $110,337 on AWS AgentCore. Those unit economics make startups switch providers.
For 640,000+ customers already on the platform, adopting AI is a feature toggle, not a migration. Agents, inference, data, and compute on one system, billed on one invoice, eliminating the egress taxes that come from stitching across vendors.
Stand Insurance Built a World Model That Simulates How Fire Spreads From a Specific Tree to a Specific House
The same architecture Silver described at Ineffable, applied to property insurance.
Stand Insurance launched a "world model for the built environment": a physics-based digital twin that simulates fire, hurricane, severe storms, flood, and earthquake at individual property level. Validated against the Palisades fire: 78% accuracy versus 44% for location-only models. The platform prices risk using the simulation, then reruns with mitigation applied, showing exactly how specific actions change the premium.
Why this matters:
This is a world model in the same sense David Silver uses the term at Ineffable (Issue #104). It predicts the next state of an environment through simulation, not by pattern-matching on historical data. The 78% vs 44% accuracy gap on the Palisades is the difference between physics and statistics.
The mitigation pricing loop is the commercial insight. Removing trees: risk down 22%, premium down $1,980. Replacing mulch with gravel: risk down 15%, premium down $1,350. Every dollar spent on resilience saves $13. That creates a market for mitigation products that previously had no way to prove their value.
Chatbots and coding agents are increasingly no longer the only game in town.
Anthropic Explores DRAM-less Inference Chips From Fractile
The memory crisis is now driving chip architecture decisions.
Anthropic is in early talks with London-based Fractile about purchasing its inference accelerators, according to The Information. Fractile's chip co-locates memory and compute on the same die using SRAM, eliminating the off-chip DRAM bottleneck. Founded by Oxford PhD Walter Goodwin, backed by the NATO Innovation Fund. Raising $200 million at a $1 billion+ valuation. Commercial chips expected around 2027.
Why this matters:
DRAM costs $9.71 per gigabyte in 2026, up 158% from 2025. A chip that eliminates DRAM entirely becomes dramatically more attractive when the component it removes is the one the entire industry cannot get enough of.
This would make Fractile Anthropic's fourth chip supplier alongside NVIDIA, Google TPUs, and Amazon Trainium. If SRAM-based inference delivers on the simulated claims (100x faster, 10x cheaper), it reshapes the cost structure dragging on Anthropic's margins despite $30 billion in revenue.
Fractile is London-based, Oxford-founded, NATO-backed. Last issue's top story was Ineffable, another UK AI company receiving sovereign capital. The UK keeps producing AI silicon and AI research companies that global buyers want.
Microsoft Orders 66,000 Next-Gen GPUs Through Nscale
The world's largest cloud provider is locking in post-Blackwell silicon before Blackwell is fully deployed.
Nscale announced an expansion of its Microsoft collaboration: 66,000+ NVIDIA Vera Rubin NVL72 GPUs at Start Campus in Sines, Portugal, starting late 2027. The deal builds on 12,600 Blackwell Ultra GPUs already deploying in the first building. Nscale is investing an additional €230 million in shared infrastructure and €465 million in a second 200MW building, roughly €695 million in new capital into a single site. Start Campus is fully permitted for 1.2GW across six buildings, powered by 100% renewable energy with ocean cooling (target PUE 1.1, WUE zero). Across all sites, Nscale has now committed over 116,000 GB300 GPUs plus 96,000 Vera Rubin GPUs to Microsoft in Portugal, Norway, the UK, and the US.
Why this matters:
Nscale's catalogue arc: Issue #101 OpenAI cancelled Stargate UK. Issue #102 Nscale landed Microsoft, Google, BT. Issue #104 EE first Iceland deployment in three months. Now: 66,000 next-gen GPUs plus €695 million in new site investment for Microsoft. Losing your anchor tenant is survivable if you replace them with the world's largest cloud provider and then keep expanding.
Microsoft is ordering Vera Rubin while most of the market is still deploying Blackwell. GPU generations sell out before they ship.
Microsoft locking in 96,000 next-gen GPUs through Nscale alone (66,000 Portugal, 30,000 Norway) tells you the hyperscalers are already competing for post-Blackwell allocation, and they are using neoclouds as procurement vehicles to secure it.
Peter Thiel Leads $140 Million Into Wave-Powered Data Centres That Sail Themselves to the Open Ocean
Build where nobody can tell you no.
Peter Thiel is leading a $140 million investment in Panthalassa, a startup building 85-metre solid-steel data centre structures that sit below the ocean surface, powered by wave energy and cooled by seawater. The nodes propel themselves using hull shape alone, no engine. AI chips receive queries via SpaceX Starlink. Electricity generated on-site, never transmitted to shore. Valued at close to $1 billion. Other investors: Marc Benioff, Max Levchin, John Doerr. Commercial deployments targeted for next year.
Why this matters:
Last issue covered Google building $15 billion of capacity in India because India said yes while Virginia and Utah said no (Issue #104). Panthalassa takes the next step: leave land entirely. No zoning, no grid, no utility. The facility generates its own power, cools itself with the ocean, and communicates via satellite.
Thiel funded the Seasteading Institute to build communities in international waters beyond sovereign jurisdiction. Panthalassa applies that thesis to AI compute.
The open ocean, connected via Starlink, powered by waves, is the most literal expression of the idea that infrastructure will route around terrestrial constraints.
The Rundown
Count the workarounds in this issue.
Anthropic needed inference capacity it could not build fast enough. It rented the cluster its competitor built and failed to use. Fractile is designing chips without the memory component the industry cannot source. Panthalassa is building data centres in the ocean because the permitting process on land has stalled. Meta is cutting 8,000 jobs because infrastructure spending has outgrown the headcount budget. NVIDIA is taking equity in its own customers because selling GPUs alone no longer captures enough of the value chain.
None of these are normal scaling decisions - normal scaling is build more of what works.
Yet every story this week is a company that hit a wall with what works and went around it.
That pattern has a name in other industries. When oil companies started drilling horizontally because the easy vertical wells ran dry, the industry called it unconventional extraction. The resource was still there, but the conventional approach to reaching it had exhausted itself.
By that logic, the AI infrastructure market just entered its unconventional phase.
And the resource is compute.
The conventional methods of reaching it, building data centres on land, buying DRAM from two suppliers, hiring people to manage the buildout, and selling GPUs as a neutral vendor, are all hitting limits at the same time. The companies in this issue are not solving those limits, but routing around them in real time.
See you next week.