Compute & Infrastructure

Jensen Huang's acknowledgement that Nvidia now holds 'zero percent' market share in China is one of the most consequential admissions in recent semiconductor history. China represented a meaningful and growing revenue base for Nvidia's data centre business before successive rounds of U.S. export controls — targeting first the A100, then the H100, and eventually the downgraded H20 — systematically closed the door. Huang's characterisation of the policy as having 'already largely backfired' reflects a strategic reality: the controls have not slowed Chinese AI development but have instead accelerated domestic alternatives from Huawei (Ascend 910C), Cambricon, and Biren, while permanently transferring that market share away from U.S. firms. Tom's Hardware

The strategic damage is compounding. Chinese hyperscalers and model developers, previously dependent on Nvidia's software ecosystem (CUDA) as much as its hardware, are now being forced to build CUDA-alternative toolchains — a transition that, once complete, removes a durable switching cost that has been Nvidia's deepest moat. The medium-term risk is not just lost China revenue (estimated by analysts at $12–15 billion annually at peak), but the emergence of a parallel AI hardware ecosystem that could eventually compete in third markets.

The shift from training-dominated to inference-dominated AI workloads is creating the most credible competitive opening against Nvidia since the GPU became the default AI accelerator. As The Register reports, inference workloads have fundamentally different hardware requirements — lower precision arithmetic, latency sensitivity over throughput maximisation, and a premium on memory bandwidth efficiency over raw FLOP counts — that play to the architectural strengths of purpose-built inference accelerators. The Register

The Anthropic-Fractile discussions are the sharpest concrete signal of this dynamic. Fractile's SRAM-based architecture eliminates external DRAM entirely, reducing both latency and dependence on the HBM memory supply chain that currently flows almost exclusively through SK Hynix and Samsung at the high end. If Anthropic converts early discussions into a procurement commitment, it would validate SRAM-centric inference accelerators commercially and potentially trigger similar conversations at other labs. Tom's Hardware This is still an early-stage discussion, not a confirmed contract — the gap between vendor conversations and volume procurement in this market is significant.

Reports that OpenAI missed both revenue and user targets in recent periods, and faces internal concern about meeting data centre spend commitments, are significant not just as a company-specific financial story but as a stress test of the broader model by which AI infrastructure is being financed. The Stargate programme — a joint venture with SoftBank, Oracle, and others targeting $500 billion in U.S. AI infrastructure investment — was structured around assumptions of rapid OpenAI revenue growth that would justify the capital deployment. If those assumptions are slipping, the renegotiation of capacity commitments becomes a real operational risk for data centre developers and power infrastructure planners who have already committed resources. Data Centre Dynamics

This connects directly to the credit market signal from CreditSights: after $300 billion in AI-linked debt issuance across the credit spectrum, investor selectivity is increasing. Bloomberg Infrastructure developers who assumed continuous access to cheap debt financing for data centre construction are now facing a more discriminating market that will require stronger offtake agreements and demonstrable revenue visibility — exactly the thing that OpenAI's reported shortfalls undermine.

Nvidia's decision to accelerate end-of-life timelines for older DDR4-based Jetson embedded AI modules — driven by DRAM supply constraints rather than a planned product transition — illustrates how memory shortages are propagating beyond high-end HBM for training clusters and into the broader AI hardware ecosystem. While the Jetson discontinuations primarily affect edge AI and robotics deployments rather than data centre infrastructure, the mechanism is the same: memory allocation is being prioritised toward higher-margin, higher-volume products, leaving legacy platforms unsupported earlier than customers planned. Tom's Hardware

This is the context in which Fractile's SRAM-based inference chip architecture carries strategic weight beyond its performance characteristics. SRAM eliminates dependence on the commodity and specialty DRAM supply chains that are currently constrained — a genuine supply chain diversification argument, not merely a technical one. Memory supply concentration at SK Hynix and Samsung for HBM, combined with broader DRAM tightness, represents one of the most underappreciated chokepoints in the current AI infrastructure buildout.