Compute & Infrastructure

Analysts now project that Chinese-designed AI accelerators will supply approximately 90% of domestic AI processor demand in 2026, with Huawei's Ascend line and Cambricon identified as the primary beneficiaries of the structural shift away from NVIDIA and AMD, according to Tom's Hardware. These are analyst projections, not confirmed procurement figures, and should be treated as directional rather than precise — the actual mix will depend on software stack maturity and yield rates at SMIC, which remains constrained to less advanced nodes than TSMC.

The strategic implication is significant regardless of the exact number: US export controls, rather than suppressing Chinese AI development, appear to have accelerated a domestication of the supply chain that will be difficult to reverse. If Huawei's Ascend 910C and successor chips achieve adequate performance-per-watt for large model training, the leverage that NVIDIA currently holds over global AI infrastructure will be structurally reduced in the world's second-largest AI market. The remaining question is whether Chinese cloud providers and model labs will voluntarily use domestic chips for frontier training or only for inference and edge deployment.

Cerebras Systems announced a new inference computer it claims delivers a wider speed advantage over NVIDIA hardware, building on its wafer-scale chip architecture that prioritises low-latency token generation over raw training throughput, per Bloomberg. No independent benchmark data was provided in the announcement — a pattern consistent with prior Cerebras marketing claims that later proved difficult to replicate at scale across diverse model architectures.

Groq's situation is more complex. The LPU inference chip maker raised $350 million but at a valuation that Data Center Dynamics reports has nearly halved — with NVIDIA's participation in the round described only as 'planned,' not confirmed. A lower valuation post-NVIDIA licensing deal suggests the market interpreted the license as a signal of competitive limitation rather than validation. Together, Cerebras and Groq illustrate that the inference chip opportunity is real but monetisation paths are narrowing as NVIDIA invests in its own inference optimisation software and hyperscalers develop in-house silicon.

AMD claimed its 2026 rack-scale AI solution delivers 4x better energy efficiency compared to its 2024 platform, and stated the company is ahead of schedule on a 20x efficiency improvement target by 2030, according to Tom's Hardware. Critically, the company did not release actual benchmark results to substantiate the 4x figure, making independent verification impossible at this stage. The claim is best treated as a roadmap signal rather than a confirmed specification.

The strategic context for this claim matters: data centre operators are under real grid capacity pressure, and energy efficiency gains at the chip and rack level are one of the few levers they can pull without waiting for new utility connections. If AMD's efficiency trajectory is directionally accurate — even if the exact multiple is optimistic — it changes the calculus for operators choosing between NVIDIA and AMD clusters when power budgets are fixed. Separately, the Sunrun-Voltus arrangement to aggregate residential solar and storage capacity for AI data centres illustrates how operators are reaching into unconventional power supply channels, per Data Center Dynamics. This is a workaround for grid constraints, not a solution to them.

Robeco's Asia fixed income head flagged that the cumulative AI capital expenditure wave — estimated at roughly $1 trillion — is large enough to materially expand corporate bond supply and sustain elevated volatility in fixed-income markets, per Bloomberg. This is an important infrastructure financing signal: as hyperscalers exhaust internally generated cash flows, they will increasingly access debt markets, competing with sovereign and corporate issuers for investor appetite. The spillover into fixed income is underappreciated by analysts who track AI capex primarily through equity lenses.

On the equity side, Rebellions — a South Korean AI chip startup — is in active IPO preparations targeting a listing on Korea's main exchange, with its CFO confirming the plans publicly at the Seoul AI Summit, per Bloomberg. A Rebellions listing would be a test of whether sovereign AI chip ambitions in mid-tier markets can attract public market capital, particularly against the backdrop of the broader semiconductor selloff. South Korea's government has strategic incentives to support a domestic AI chip champion as a hedge against Samsung and SK Hynix's dependence on NVIDIA's ecosystem for HBM demand.