Capital & Industrial Strategy

Amazon is injecting a confirmed $5 billion into Anthropic immediately, with the WSJ and Bloomberg reporting an option to deploy up to $20–25 billion more over time, bringing the potential total Amazon commitment to $25 billion. In return, Anthropic has committed to spending over $100 billion on AWS infrastructure over the next decade, including securing up to 5 gigawatts of compute capacity. The deal was confirmed by both parties and reported across the FT, Bloomberg, WSJ, TechCrunch, and CNBC — the core terms are closed, though the incremental $20 billion is conditional.

The strategic logic is layered. Amazon is not simply buying equity upside in Anthropic; it is manufacturing a captive hyperscaler customer at unprecedented scale. The $100 billion cloud commitment insulates AWS from losing Anthropic's workloads to Azure or GCP, which matters acutely given that Microsoft's OpenAI relationship and Google's Gemini infrastructure have both deepened this year. For Anthropic, the deal resolves a near-term operational crisis — the company suffered notable outages earlier in 2026 — by guaranteeing access to compute at a scale it could not independently procure. The risk is structural dependency: Anthropic's cost base is now explicitly tethered to AWS pricing, and any deterioration in that relationship would be existential. Regulators in the EU and UK, who have already scrutinised prior Amazon-Anthropic arrangements, will almost certainly re-examine this expanded tie-up.

The FT reports that Jeff Bezos's secretive AI lab, operating under the codename Project Prometheus, is close to closing a $10 billion funding round that would value the company at approximately $38 billion. Bloomberg confirmed the FT reporting. The lab is focused on AI models capable of understanding the physical world — language distinct from pure language modelling and clearly targeting robotics, industrial automation, and embodied AI applications. The deal is described as near-final but has not yet closed; valuation and round size should be treated as announced intention, not confirmed terms.

The $38 billion pre-revenue valuation reflects a broader market dynamic: capital is pricing in winner-take-most outcomes in physical-world AI before any revenue benchmark exists. This is structurally similar to the early rounds of OpenAI and Anthropic, but at a higher absolute valuation, suggesting investor risk appetite for frontier AI has not compressed despite macro uncertainty. Bezos's personal brand and his proximity to Amazon's robotics and logistics infrastructure — a potential distribution channel for physical-world models — are almost certainly part of the valuation justification.

Google is expected to announce a new generation of inference-focused Tensor Processing Units this week, per Bloomberg reporting citing specific product details on the TPU roadmap. Simultaneously, Reuters and CNBC report that Marvell is in active deal talks with Google to co-develop two custom AI chips — news that sent Marvell shares up while Broadcom shares fell, reflecting market interpretation that Google may be shifting custom silicon co-development work toward Marvell. Marvell received a $2 billion Nvidia investment in March, making it a strategically contested asset.

The inference-chip focus is analytically significant. Training compute is largely Nvidia-dominated and difficult to displace given CUDA's ecosystem lock-in. Inference is a structurally different market: higher volume, more cost-sensitive, and less dependent on the software ecosystem that entrenches Nvidia in training. Google deploying inference-optimised TPUs positions it to undercut Nvidia on the workloads that will dominate enterprise AI spend over the next three to five years — the running of deployed models at scale. Morgan Stanley separately notes that agentic AI is already widening chip spending beyond GPUs into CPUs, reinforcing the thesis that inference and agent execution hardware is the next spending battleground.

Three converging data points illustrate the physical limits of the current AI buildout. First, Semafor reports that Apple Mac Mini computers are effectively out of stock globally because the machine has become the most cost-effective platform for locally hosted AI agents — a granular demand signal indicating enterprise and developer adoption of agent infrastructure is outpacing supply even at the commodity hardware level. Second, Reuters reports energy constraints are becoming a binding limit on Big Tech's AI profit projections, with power availability now a gating factor for data centre expansion. PGIM's real estate leadership separately noted that the AI data centre boom is testing buildout limits, a view from the capital markets side of infrastructure financing.

Third, ByteDance's profit fell more than 70% year-on-year as the company leaned aggressively into AI investment — a concrete P&L illustration of the margin destruction underway at hyperscale AI spenders. For investors, ByteDance's numbers are a leading indicator: companies choosing to compete at the frontier are accepting near-term profitability destruction for strategic positioning, and this dynamic will increasingly surface in Western public company earnings as well. The Fermi nuclear startup's sudden CEO and CFO departures — the company co-founded by former US Energy Secretary Rick Perry was building an AI campus in Texas — adds a cautionary note on the nuclear-for-AI thesis, where execution risk is proving significant despite strong strategic logic.