AI Sector Funding Concentration: Winner-Take-All Dynamics in 2026

As discussed in relation to China's AI unicorn cohort (Zhipu AI, MiniMax, Moonshot AI — now including two January 2026 Hong Kong IPOs, covered in depth in a related fact), a small handful of foundation-model companies absorb a disproportionate share of the sector's total funding, a concentration curve that superficially resembles the OpenAI/Anthropic-dominated shape of the US AI funding landscape. The mechanism producing that concentration differs meaningfully, however: US concentration reflects private capital markets converging on the small number of teams with genuine technical differentiation and compute access, largely independent of government direction, while Chinese concentration is substantially shaped by industrial policy actively identifying and channeling capital toward a specific shortlist of companies — several of which have now been fast-tracked to a public listing as a matter of apparent policy preference.

The Nvidia H20 episode is the clearest possible illustration of how unstable compute access has become as a competitive variable, and it is worth tracing in its actual documented sequence rather than as a general claim. Nvidia's H20 chip was specifically designed to comply with the US government's October 2023 export controls on advanced semiconductors to China — a lower-performance product built deliberately to stay on the legal side of the restriction. Nvidia told investors it expected to sell more than one million H20 units to China in 2024, generating roughly $12 billion in revenue. Then, in April 2025, the US Commerce Department reversed course and declared the H20 — the same chip previously deemed compliant, and the same chip that had powered the DeepSeek AI model's widely reported breakthrough — newly non-compliant with export controls. In July 2025, Commerce reversed itself again, agreeing to grant Nvidia export licenses for the H20 after all.

What happened next is the part that complicates any simple narrative: despite the US reopening the door, Nvidia never actually sold H20 chips into China after this sequence, because Chinese authorities told domestic AI companies not to purchase them, citing security concerns of their own — and Nvidia subsequently stopped manufacturing the chip and took a $4.5 billion inventory and purchase-obligation charge in the first quarter of its 2026 fiscal year as a direct result. This is not a story of the US unilaterally restricting China's compute access; it is a story of both governments independently constraining the same transaction from opposite sides within the same twelve-month window, leaving Chinese AI companies with less access to this specific chip than either government's stated policy alone would have predicted.

The H20 episode suggests China's AI sector concentration dynamics are being shaped by a genuinely two-sided compute-access constraint, not a purely external one — Chinese regulators are willing to discourage even legally available foreign chips on security grounds, which plausibly accelerates domestic compute self-sufficiency efforts (a core "new quality productive forces" priority) at the cost of near-term performance for companies that might otherwise have preferred the foreign chip. Companies with both political backing and access to whatever advanced or domestically-produced compute capacity exists have a structural advantage that is partially independent of their underlying technical talent or private fundraising success — a dynamic with no close parallel in the US market, where compute access, while expensive, is not simultaneously export-controlled from outside and discouraged from inside.

China's response to the compute constraint described above has centered substantially on Huawei's Ascend AI chip line and domestic foundry capacity, principally SMIC (Semiconductor Manufacturing International Corporation), as the two most advanced domestic alternatives to foreign advanced AI chips. Both have made genuine progress and receive priority state support and procurement preference, but both also remain, by most independent technical assessments, one to several process-node generations behind the most advanced foreign alternatives in raw performance and manufacturing yield — meaning a Chinese AI company relying primarily on domestic chips today generally accepts a real performance and cost-efficiency disadvantage relative to a hypothetical unconstrained-access competitor, a gap that state subsidy can help offset commercially but cannot fully eliminate on pure technical performance grounds in the near term. This gap, not simply the existence of a domestic alternative, is what makes the ongoing back-and-forth over foreign chip access (illustrated by the H20 saga) commercially consequential rather than a purely symbolic dispute.

Compute access and talent access compound each other in a way worth stating explicitly: a company with constrained compute access is also a less attractive place for a top AI researcher to do their best work, since research productivity in this field is directly bottlenecked by available training compute — meaning the compute constraint discussed throughout this fact indirectly worsens the talent-competition dynamic discussed in a related fact on the broader AI Tiger cohort, compounding rather than operating independently of it. A foreign investor evaluating a mid-tier or emerging Chinese AI company, rather than the top-tier cohort with strong state and strategic backing, should specifically weight this compounding effect — a company with both weaker compute access and weaker resulting ability to attract frontier research talent faces a double, self-reinforcing disadvantage relative to the sector's leaders.