Inference Wars, Capital Concentration, and the Agentic Land Grab

Three converging data points this week crystallise a structural shift: xAI's 220,000-GPU Colossus 1 supercluster failed as a training system due to mixed-architecture inefficiencies and has been leased in its entirety to Anthropic for inference workloads; Cerebras — whose entire product thesis is inference-optimised wafer-scale silicon — debuted at $67 billion; and SambaNova's CEO used Cerebras' IPO moment to publicly frame inference cost-per-token as the real competitive frontier. The DRAM ETF reaching $10 billion in assets at record pace reflects institutional money repositioning toward memory bandwidth, the binding constraint in inference at scale, not training silicon.

The commercial logic extends beyond hardware. OpenAI's Plaid integration, Amazon's full deployment of Alexa as its core search interface, and OpenAI's organisational pivot to agents are all inference-layer events — they represent the deployment of already-trained models into real-world execution contexts at massive scale. Training compute determines what models can do; inference infrastructure determines who captures the economic value from doing it. Capital allocators tracking inference cost-per-token across major providers now have a more reliable leading indicator than training benchmark comparisons.

China's voluntary embargo on H200 purchases — disclosed by Trump in direct bilateral talks — is analytically distinct from US export controls: it is demand-side sovereign policy designed to accelerate domestic silicon adoption from Huawei and emerging Chinese fabless players. The constraint on Nvidia's China revenue is now self-reinforcing on both sides, eliminating any near-term scenario where commercial pressure reopens the market. Simultaneously, the FTC's antitrust probe into Arm's licensing practices introduces regulatory risk into the dominant non-x86 architecture underlying virtually every mobile chip and a rapidly growing share of data centre inference silicon — an adverse outcome would restructure the cost base for every custom silicon programme at AWS, Google, Microsoft, and Apple.

Samsung's 18-day strike and pre-emptive production throttling add a supply-side shock to an already bifurcated landscape. HBM3E memory — the bandwidth-critical component for Nvidia H100 and H200 systems — is already in constrained supply from a three-player global market. Any reduction in Samsung's output propagates directly into GPU system availability with no short-term substitute. The workforce scarcity flagged by SEMI compounds this: the US is simultaneously attempting to scale domestic semiconductor production while restricting the immigration pathways that historically supplemented domestic STEM pipelines, extending fab ramp timelines beyond current projections regardless of CHIPS Act capital commitments.

The most consequential moves this week were not capability breakthroughs but integration land grabs. OpenAI's Plaid deal grants ChatGPT structured access to 12,000 financial institutions, moving it from discussing finances in the abstract to analysing actual spending and — as the agentic roadmap implies — eventually executing transactions. Amazon's full production deployment of Alexa as the default Amazon.com search interface collapses conversational AI into the world's largest product discovery engine, directly threatening both Google Shopping and Amazon's own high-margin keyword advertising business. Microsoft's cancellation of Claude Code licences signals the transition from exploratory multi-vendor experimentation to platform consolidation, with major enterprises hardening their stack around proprietary or strategically aligned tooling.

Meta's encrypted Incognito Chat represents the countermove: architectural privacy as a competitive differentiator that unlocks regulated professional markets — legal, medical, financial — that are structurally off-limits to data-integrating AI assistants regardless of capability. The tension is direct. OpenAI is expanding its data integration surface to build agentic utility; Meta is contracting its data surface to build professional trust. Both are rational strategies for different segments, and the outcome will depend on whether enterprise buyers in regulated industries trust Meta's privacy claims despite its advertising history. Independent cryptographic audit of the implementation is the outstanding verification requirement that will determine whether this is genuine differentiation or marketing positioning.