China's GPU Developers Post First Profits Amid Memory Cost Pressures

Milestone Quarter for Domestic Designers

Three Chinese GPU developers - Moore Threads, MetaX and Iluvatar CoreX - reported positive operating income for the three months ending September 2026, marking the first time any of the country's pure-play AI accelerator companies have escaped losses since their founding. The collective swing came as data-centre operators in China, cut off from Nvidia's latest H-series and A-series accelerators by US export restrictions, placed orders for domestically designed alternatives at volumes large enough to offset the steep research and tooling costs that have held the sector underwater.

Moore Threads, which shipped approximately 18,000 units of its MTT S80 inference card during the quarter, logged an operating profit of ¥47 million (US$6.5 million) on revenue of ¥890 million. MetaX, whose C500 training accelerator has found traction among state-owned cloud providers, posted ¥31 million in profit on ¥620 million of sales. Iluvatar CoreX, backed by Tencent Holdings and focused on edge inference, reported a narrower ¥12 million gain. All three firms had accumulated combined losses exceeding ¥4 billion over the previous eight quarters.

The turnaround reflects a structural shift in procurement behaviour. Since mid-2025, when the US Department of Commerce tightened export thresholds to restrict any chip capable of more than 300 teraflops of dense matrix performance, Chinese hyperscalers and regional cloud operators have had little choice but to design workloads around locally available silicon. Interviews with three procurement managers at tier-two data-centre operators, conducted in October, confirmed that lead times for domestic GPUs have compressed to four to six weeks, compared with indefinite waits - or outright unavailability - for embargo-affected hardware.

Memory Costs Threaten Margin Sustainability

Yet the profitability milestone coincides with a sharp deterioration in a critical input market. High-bandwidth memory, the specialised DRAM that sits alongside GPU dies and feeds them training data, has climbed 63 per cent in contract price since January 2026, according to TrendForce. HBM3 modules, which all three Chinese vendors integrate into their flagship products, are now trading at approximately $950 per 24-gigabyte stack, up from $580 a year earlier.

The increase stems from a confluence of supply and demand pressures. SK hynix, Samsung and Micron Technology together control more than 95 per cent of HBM production capacity, and all three have prioritised shipments to Nvidia, AMD and other established customers under multi-year agreements. At the same time, Chinese buyers - operating outside those long-term contracts - must compete in the spot market or accept shorter-term pricing that resets quarterly. One senior sourcing executive at a Shenzhen-based AI hardware firm, speaking on condition of anonymity, estimated that his company's HBM costs per unit rose by 58 per cent between Q2 and Q3 2026, erasing roughly half the gross-margin improvement achieved through higher sales volumes.

Compounding the challenge, none of the Chinese GPU developers has secured preferential allocation agreements with the three leading memory makers. Changxin Memory Technologies, China's most advanced domestic DRAM producer, is still ramping pilot production of HBM2E - a generation behind the HBM3 standard now required for competitive inference and training performance. Industry observers do not expect Changxin to reach meaningful HBM3 output until the second half of 2027 at the earliest, leaving Chinese designers dependent on imports for the next four to six quarters.

Competitive Dynamics and Design Trade-offs

The three profitable firms are pursuing divergent strategies to navigate the cost squeeze. Moore Threads has focused on inference accelerators for large-language-model deployment, a segment that tolerates slightly older HBM2E memory and allows the company to source from Changxin's existing lines at a discount to spot HBM3 prices. MetaX, by contrast, is targeting training clusters for frontier-model research, a use case that demands maximum memory bandwidth and forces reliance on imported HBM3. Iluvatar CoreX has carved out a niche in edge AI, where lower per-chip memory requirements reduce exposure to HBM price swings but also limit addressable revenue per unit.

The divergence reflects broader uncertainty about where sustainable profit pools will emerge. Training workloads command higher prices per chip - MetaX's C500 lists at approximately ¥78,000, compared with ¥52,000 for Moore Threads' S80 - but carry thinner margins once memory costs are factored in. Inference deployments generate steadier, higher-volume demand, yet face potential commoditisation as software optimisation reduces the performance gap between domestic and embargoed hardware.

A further complication is the pace of architectural iteration. Nvidia's roadmap cadence, historically annual, has slowed under export restrictions, giving Chinese competitors a longer window to close performance deficits. At the same time, the US firms' lead in software ecosystems - CUDA remains the de facto standard for model training - means that even functionally comparable Chinese silicon requires costly middleware and developer outreach to win design wins. Moore Threads, for instance, has allocated roughly 40 per cent of its engineering budget to MUSA, its CUDA-compatible runtime, rather than to next-generation chip design.

Forward-Looking Indicators

Several factors will determine whether Q3 2026 represents an inflection point or a brief interlude. On the supply side, any easing of US export policy - whether through revised Commerce Department rules or bilateral negotiation - would reintroduce Nvidia and AMD as direct competitors, likely compressing prices and market share for domestic players. Conversely, further tightening, such as restrictions on advanced packaging services provided by Taiwan Semiconductor Manufacturing Company, could disrupt Chinese production lines and delay product launches.

On the demand side, China's data-centre capital expenditure growth is projected to decelerate in 2027 as earlier AI infrastructure buildouts mature and operators shift spending toward inference optimisation and energy efficiency. A September 2026 survey by the China Academy of Information and Communications Technology found that 68 per cent of cloud providers plan to hold GPU procurement budgets flat or reduce them in the coming fiscal year, a marked shift from the double-digit growth seen in 2025 and early 2026.

Memory supply, meanwhile, hinges on Changxin's ability to scale HBM3 production and on whether Chinese chipmakers can negotiate volume commitments that justify dedicated capacity from Samsung or SK hynix. Neither outcome is certain. Changxin's HBM3 pilot line has encountered yield issues related to through-silicon via alignment, according to two engineers familiar with the programme, and the company has not yet published a formal production timeline. Samsung and SK hynix, for their part, face political and compliance risk in signing long-term deals with Chinese GPU developers, given the possibility of secondary sanctions or entity-list designations.

Margin Outlook and Strategic Responses

Analysts at Bernstein Research estimate that if HBM3 prices remain at current levels, gross margins for Moore Threads, MetaX and Iluvatar CoreX will compress by 8 to 12 percentage points over the next four quarters, potentially returning all three firms to operating losses by mid-2027. The forecast assumes flat average selling prices and modest volume growth, in line with the capital-expenditure outlook described above.

In response, the three companies are exploring vertical integration. Moore Threads has invested in a joint venture with a Jiangsu-based packaging house to develop in-house 2.5D interposer capabilities, aiming to reduce reliance on TSMC's CoWoS service and lower per-unit costs by approximately 15 per cent. MetaX is in discussions with Changxin to co-develop a custom HBM variant optimised for training workloads, potentially sacrificing some bandwidth for lower cost. Iluvatar CoreX, with its edge focus, is evaluating LPDDR5 memory as a substitute for HBM in lower-performance SKUs, though such a shift would require software re-optimisation and may limit compatibility with existing customer deployments.

Government support remains a variable. The National Integrated Circuit Industry Investment Fund, China's state-backed semiconductor vehicle, has historically provided grants and equity injections to sustain loss-making chipmakers through development cycles. Whether it will continue that support now that firms have crossed into profitability - and whether any such support would trigger countervailing duties or other trade responses - is unclear.

The third-quarter results demonstrate that Chinese GPU developers can achieve profitability under the current policy and market configuration. Sustaining those gains, however, will require navigating input-cost inflation, demand moderation and the ever-present risk of further export restrictions. For now, the sector's financial health is as much a function of geopolitics and memory markets as of chip design itself.