Today's News
Data Gravity by Chris Zeoli
Chris Zeoli's analysis of GPU-backed lending reveals a fundamental bifurcation in AI infrastructure financing. Companies with investment-grade customer contracts (hyperscalers) borrow at ~7% via structured SPVs, while neoclouds without anchor tenants face 12–16% rates—an $80M annual cost difference on $1B of equipment that determines pricing power and survival. CoreWeave exemplifies the dynamic: its cheapest loans cost 7% (backed by top-tier contracts), while its unsecured bonds trade at junk spreads (79 cents on the dollar, 11.9% default insurance) despite averaging only 8.4% across its $35.6B debt portfolio. The market is explicitly pricing offtaker credit and contract tenor, not GPU residual value. Spreads have compressed dramatically from 2023's ~870bps to 2026's 225bps for investment-grade structures, saving ~$550M annually on deals like CoreWeave's $8.5B DDTL 4.0. The template—pairing GPUs with assigned customer contracts in bankruptcy-remote SPVs—has become the standard across Lambda, Nebius, and others, with Nvidia and Google increasingly backstopping tail risk through vendor credit support programs targeting $500B in financing capacity.
James Titcomb / Telegraph via Techmeme
Microsoft's withdrawal from Nscale's Essex facility—which was to house 23,000 Nvidia chips—exposes a critical constraint: grid power won't arrive until the 2030s. The exit threatens Nscale's $35B IPO ambitions, with investors now pressuring for a $16–19B valuation. This signals that infrastructure buildout timelines, not just capital or chip availability, are becoming material blockers for hyperscaler expansion plans in key markets.
Anthropic via Noahpinion
Empirical data presented by Ramez Naam and Noam Brown reveals that adding more parallel AI agents produces sub-logarithmic intelligence gains on complex tasks like formal math proofs and Navier-Stokes problems. Expanding from ten to one hundred agents yields almost no additional capability, directly contradicting theoretical models predicting that millions of agents could sustain recursive self-improvement. Brown's work on OpenAI's breakthrough emphasized that the 10,000-agent system's success depended primarily on underlying model quality rather than sheer parallelization, challenging the narrative that agent count alone drives capability breakthroughs needed for AI self-improvement loops. This finding has significant implications for infrastructure planning: it suggests that the "country of geniuses in a datacenter" faces fundamental limits that cannot be overcome by simply deploying more compute.
SemiAnalysis
SemiAnalysis satellite imagery and construction tracking reveal 24.3GW of AI datacenters in advanced stages across 40 major campuses, with most sites over 50% complete and cooling/electrical systems already installed. Year-to-date, slightly under $1 trillion in binding and non-cancellable datacenter leases have been signed—a leading indicator that outpaces actual buildout by 12–18 months and directly supports the 2027–28 acceleration thesis. The pool of buyers has broadened dramatically, with Meta, AWS, Microsoft, Google, Nvidia, AMD, Anthropic, and OpenAI all competing for GW-scale capacity. This concrete evidence of on-time delivery at scale (Crusoe Abilene, AWS Indiana, QTS Georgia) directly contradicts skeptics like Altimeter's Brad Gerstner, who claimed only 25GW would be built in 2027 versus SemiAnalysis's +50GW forecast. Hyperscalers have also innovated on design—Google modularized electrical architecture, Meta deployed tent designs, and AWS launched "SAMDC" rapid design with construction timelines under 10 months, reducing MEP labor by up to 63%. Behind-the-meter power generation is solving grid constraints that would otherwise limit 2027 buildout.
Jack Clark via Import AI
At the IMF's Per Jacobsson Lecture in Bangkok, Jack Clark delivered the argument that AI systems have begun operating autonomously at the frontier of multiple scientific disciplines simultaneously. Concrete evidence includes FrontierMath improving from 2% to 94% accuracy on previously unsolved problems and GPT-6-Astra scoring 98% on TheoPhysEval (theoretical physics), even correcting expert answers. Clark positions AI-driven scientific capability as the inflection point triggering exponential economic growth and argues policymakers must immediately build scientific infrastructure—labs, robotics, equipment—to capture the benefits. This framing matters for infrastructure planning: if frontier science becomes the binding constraint on AI progress rather than compute, the implications for datacenter buildout and capital allocation shift materially.
US Hyperscaler Capex Faces Critical Dependency on Chinese Rare Earths Despite Eight Years of Tariff Policy
Azeem Azhar, Exponential View
Despite eight years of tariff policy since 2018, China retains gatekeeper control over essential inputs for AI infrastructure. China supplies 70% of US rare earths and yttrium; April 2025 export restrictions dropped yttrium shipments from 38 tonnes monthly to 2 tonnes, causing European yttrium prices to rise 45-fold. With $800 billion in data center capex this year relying on these components, supply-chain remapping through tariff-jumping investments (Chinese factories in Vietnam and Mexico) has proven more effective than tariff reduction. Chinese value-added in US imports dropped only from 84% to 75% over seven years, while Chinese-owned factories now ship 25% of Vietnam's exports to the US, up from 11% in 2018–19.
