Digital Colliers Daily Briefing — May 7, 2026
The compute-supply crunch that has shaped AI economics for the past year moved from subtext to headline today. Anthropic took over xAI's flagship Memphis data center in a deal estimated at roughly $5 billion a year, DeepSeek opened its first outside funding round at a $45 billion valuation backed by China's state IC fund, and SpaceX filed plans for a Texas semiconductor complex it is calling "Terafab" with a potential price tag of $119 billion. Three different answers — rented capacity, sovereign capital, vertical integration — to the same underlying problem.
1. Anthropic Absorbs Colossus 1, Lifts Claude Code Limits, and Rewrites the xAI Relationship
What happened. At its Code with Claude conference in San Francisco, Anthropic announced an agreement to use the entire compute capacity of SpaceXAI's Colossus 1 data center in Memphis — more than 300 megawatts and roughly 220,000 NVIDIA GPUs spanning H100, H200, and GB200 chips, according to the company's announcement. CTO Tom Brown said Claude inference would begin ramping on Colossus "in the next few days." Latent Space pegged the arrangement at roughly $5 billion per year, effectively turning xAI into a neocloud landlord. Elon Musk, who weeks earlier had publicly attacked Anthropic's safety posture, said xAI was comfortable leasing the cluster because training had already moved to Colossus 2, reportedly populated with around 500,000 Blackwells.
Anthropic translated the new capacity into immediate user-facing changes: Claude Code's five-hour rate limits doubled for Pro, Max, Team, and seat-based Enterprise plans; peak-hour throttling on Pro and Max was removed; and Opus API rate limits were raised substantially. Weekly caps were left untouched, which product lead Amol Avasare said was deliberate — most users hit the five-hour ceiling, not the weekly one. The Colossus deal joins Anthropic's previously announced 5 GW commitment with Amazon, 5 GW with Google and Broadcom, $30 billion in Azure capacity with Microsoft and NVIDIA, and a $50 billion Fluidstack buildout. The Information reported separately that Anthropic has committed up to $200 billion to Google TPUs, per Wired.
Why it matters. The headline is not a model — it is capacity. Dario Amodei told attendees Claude usage had grown roughly 80x against forecasts, and the company's user-facing limits only moved after a major external compute deal was struck. As Simon Willison noted in his live blog, API volume on the Anthropic platform is up 17x year-over-year. The deal also breaks the tidy narrative that each frontier lab sits atop its own vertically integrated stack: Anthropic is now running inference on hardware bought by a direct competitor.
Who is affected. Claude Code power users and Opus API customers gain real headroom today. xAI converts a depreciating asset into recurring revenue ahead of SpaceXAI's reported June IPO at a $1.25 trillion valuation. Hyperscalers Amazon, Google, and Microsoft remain Anthropic's primary backbone, but the Colossus arrangement signals that frontier labs will rent from anywhere capacity exists. Memphis residents — already organizing against gas-turbine emissions at the site, per Wired — inherit a new operator.
What to watch next. Whether weekly caps move once Colossus inference stabilizes; whether Anthropic's "Dreaming" memory and "Outcomes" rubric features in its managed-agents platform prove defensible against open harnesses; and whether the floated joint exploration of "orbital AI compute capacity" with SpaceX produces anything beyond marketing for the IPO roadshow.
Sources:
- [HN · 459↑] Higher usage limits for Claude and a compute deal with SpaceX — Hacker News
- [AINews] Anthropic-SpaceXai's 300MW/$5B/yr deal for Colossus I, ARR growth is 8000% annualized — Latent Space
- Anthropic raises Claude Code usage limits, credits new deal with SpaceX — Ars Technica
- not much happened today — smol.ai News
- Anthropic Gets in Bed With SpaceX as the AI Race Turns Weird — Wired
- Live blog: Code w/ Claude 2026 — Simon Willison
2. DeepSeek's $45B First Round Puts China's State IC Fund at the Center
What happened. DeepSeek is in talks to raise its first-ever venture round at a valuation that has climbed from $20 billion to $45 billion in a matter of weeks, according to reporting from the Financial Times and Bloomberg relayed by TechCrunch. The round is reportedly led by the China Integrated Circuit Industry Investment Fund, Beijing's state-backed semiconductor vehicle, with Tencent and Alibaba in talks to participate. Founder Liang Wenfeng controls roughly 90% of the company and has resisted outside capital until now; the impetus, sources told the FT, is employee equity to counter aggressive poaching of DeepSeek researchers.
Why it matters. A state-led IC-fund round formalizes what has been informally true since DeepSeek's January 2025 breakout: the lab is being positioned as China's national AI champion, paired explicitly with Huawei silicon as a substitute for restricted NVIDIA hardware. The pairing — open-weight models tuned for domestic chips — is the cleanest expression yet of Beijing's strategy to route around U.S. export controls.
Who is affected. DeepSeek's researchers, who now get equity rather than just salary in the global talent war. Huawei, which gains a flagship software customer to validate its Ascend roadmap. U.S. policymakers calibrating chip controls against a counterparty that is no longer a quirky hedge-fund spinoff but a state-aligned platform. And open-weight competitors globally: DeepSeek-class models on Hugging Face have set the price-performance floor that Meta, Mistral, and Alibaba's own Qwen line are measured against.
What to watch next. Whether Tencent and Alibaba's participation closes — both have their own frontier ambitions — and at what governance terms. The forward-looking eval question is whether DeepSeek V4 Pro and successors continue to hold near-frontier benchmarks; community testing flagged today on FoodTruck Bench showed V4 Pro within roughly 3% of GPT-5.2 at about 17x lower cost, but other evaluators reported weaker results on WeirdML.
Sources:
3. SpaceX Files for a $119B "Terafab" in Grimes County
What happened. A proposal posted to the Grimes County, Texas website outlines a SpaceX plan to spend $55 billion in an initial phase — and up to $119 billion in total — on a "multi-phase, next-generation, vertically integrated semiconductor manufacturing and advanced computing fabrication facility," TechCrunch reported. Tesla is contributing resources and Intel has been pulled in as a chipmaking partner. The stated target is enough output to support 1 terawatt of chips per year, intended for AI servers, Starlink and SpaceX's proposed orbital data centers, autonomous Tesla vehicles, and Optimus-class robots. Musk noted on X that Grimes County is one of several sites under consideration.
Why it matters. If anywhere near built, Terafab would be one of the largest industrial commitments in semiconductor history — comparable in scale to TSMC's entire global capex run-rate. It is also Musk's bet on the same diagnosis Anthropic underlined this morning: chips are the binding constraint, and contracted foundry slots are not arriving fast enough. Pulling Intel into the structure offers Intel Foundry a marquee anchor customer at a moment when its leading-edge node ambitions need external validation.
Who is affected. Intel, which gains a potential megacustomer and political cover for U.S. fab investment. TSMC and Samsung, which would face their first credible vertically integrated U.S. competitor on advanced AI silicon. NVIDIA, whose pricing power assumes scarce alternatives. The combined SpaceXAI entity, valued at $1.25 trillion ahead of a June IPO, gains a long-dated supply story to pitch investors alongside the Anthropic lease and orbital-compute concept.
What to watch next. Site selection, permitting timelines, and any federal incentive package — a project of this scale will trigger CHIPS Act-era scrutiny on both subsidy and national-security grounds. Equally important: which process node Intel commits to, and whether "1 terawatt per year" survives contact with engineering reality or remains a Musk-style stretch target.
Sources:
The three stories trace a single thread: compute scarcity is now the dominant strategic variable in AI, and the industry is responding on every axis at once. Anthropic is renting its way out, even from a one-time antagonist; China is funding a vertically aligned model-and-silicon stack at the state level; and Musk's empire is attempting to build the chip supply itself, while simultaneously monetizing the cluster it just outgrew. Whichever approach scales fastest will determine more about the next 18 months of AI competition than any single model release.