Huawei just told the world it can’t keep up with its own backyard. On September 17, 2026, at Huawei Connect in Shanghai, rotating chairman Eric Xu said the company’s AI computing hardware is so oversubscribed domestically that a broad international launch isn’t on the table. Days later, on September 21-22, Tom’s Hardware detailed the scale of what Huawei is holding back: an Atlas SuperPoD cluster built from 15,488 Ascend 960 NPUs, linked by optical networking, and rated by Huawei at up to 30 FP8 exaflops and 120 FP4 exaflops.
That is a striking reversal in posture for a company that spent the past three years racing to prove its Ascend chip family could stand in for Nvidia hardware Chinese buyers can no longer get freely. Now Huawei’s problem isn’t finding customers. It’s building enough silicon to serve the ones it already has. For readers tracking the broader AI chip price surge tied to the HBM shortage, this is the flip side of the same supply crunch: demand for compute has outrun what foundries and packaging lines can physically produce, in China and everywhere else.
What Huawei Actually Announced
Xu’s comment was blunt. “Since we don’t have enough capacity to even satisfy the demand in China, we don’t have a plan to expand into the international market in a fully-fledged way,” he said, according to Reuters’ September 17 report from Huawei Connect. Read plainly, that’s a company choosing to ration its most advanced accelerators rather than chase overseas revenue it can’t currently fulfill.
This isn’t a case of Huawei lacking buyers abroad. Export-control friction with the United States already limits where advanced Chinese AI silicon can legally ship, and few Western cloud providers would risk a compliance headache to run Ascend hardware anyway. The more telling detail is the framing: domestic demand alone exceeds supply. Huawei is choosing to fully saturate the Chinese market first, and an international rollout, if it comes at all, waits until yields and packaging capacity catch up.
Tom’s Hardware’s follow-up reporting on September 21-22 filled in the technical picture Xu’s comment left out. The outlet described Huawei’s newest Atlas configuration as a direct answer to Nvidia’s rack-scale systems, built at a scale that dwarfs anything Huawei has fielded before.
Inside the Atlas SuperPoD: 15,488 Chips, 120 Exaflops
The headline number is the cluster size. Huawei’s Atlas SuperPoD, as reported by Tom’s Hardware, scales to 15,488 Ascend 960 NPUs in a single coupled system. Huawei’s own figures put that configuration at up to 30 FP8 exaflops and 120 FP4 exaflops. Those are cluster-level numbers, not per-chip specs, and they describe a fully built-out system rather than a shipping product every customer can order today.
Tom’s Hardware drew the natural comparison to Nvidia’s NVL72, the 72-GPU rack-scale platform that has become the reference design for large AI training and inference deployments. A 15,488-chip Atlas cluster is more than 200 times larger in raw chip count than a single NVL72 rack. That gap matters less than it looks, though, because Nvidia customers scale NVL72 racks out into much larger installations too, and Huawei’s own reporting acknowledged the obvious caveat: bigger doesn’t mean better per watt. Tom’s Hardware specifically noted that Huawei’s performance-per-watt is expected to trail Nvidia’s by a wide margin, which is the honest asterisk on every exaflop figure Huawei has published.
Scale-first, efficiency-second is a defensible strategy when your customers are state-directed enterprises with subsidized power and no alternative supplier. It’s a much harder sell to a commercial cloud operator paying market electricity rates and answering to shareholders about total cost of compute.
The Ascend Lineup: 910C, 910D, and the New 960
Huawei’s AI silicon roadmap now spans three distinct accelerator generations doing different jobs. The Ascend 910C, a dual-die accelerator package, has become the backbone of Huawei’s currently deployed AI clusters, with Huawei claiming up to 780 TFLOPS in BF16 precision. The 910D is the newer entrant tied to Huawei’s next-generation roadmap, positioned for domestic deployment first. The Ascend 960 is the chip Tom’s Hardware identified as the workhorse inside the largest Atlas configurations, the one pooled by the thousands to reach the SuperPoD’s headline exaflop figures.
| Chip / System | Type | Reported Performance | Current Priority |
|---|---|---|---|
| Ascend 910C | Dual-die AI accelerator | Up to 780 TFLOPS BF16 (Huawei claim) | Backbone of existing domestic clusters |
| Ascend 910D | Next-gen AI accelerator | Full specification not yet detailed publicly | Domestic rollout first |
| Ascend 960 | NPU used in Atlas SuperPoD | Pooled at scale to reach cluster totals below | Domestic rollout first |
| Atlas SuperPoD (15,488-chip) | Full rack-scale cluster | Up to 30 FP8 exaflops / 120 FP4 exaflops (Huawei claim, per Tom’s Hardware) | China-first deployment |
The dual-die design of the 910C is worth pausing on, because it’s also the source of Huawei’s biggest manufacturing headache. Packaging two dies into one accelerator package raises usable performance, but it also multiplies the number of things that can go wrong before a chip ships: die yield, interposer yield, and final package yield all compound against each other. When advanced packaging capacity is scarce, as it is across the Chinese semiconductor supply chain right now, dual-die parts eat a disproportionate share of it.
Why Optical Interconnects Matter at This Scale
Linking 15,488 accelerators into something that behaves like one coherent system is arguably harder than building the chips themselves. Tom’s Hardware reported that Huawei is using optical networking to hold the Atlas cluster together, a choice aimed squarely at the scaling penalty that shows up once electrical interconnects have to carry traffic across thousands of nodes and multiple switch layers.
Electrical signaling degrades over distance and loses bandwidth density as systems grow. Optical links sidestep both problems, at the cost of more complex transceivers and, historically, higher per-port expense. The available reporting doesn’t specify Huawei’s exact optical lane count, wavelength scheme, or end-to-end bandwidth, so those particulars remain unconfirmed. What’s clear is the direction: as AI clusters climb from hundreds of chips to tens of thousands, the network fabric connecting them becomes as strategically important as the chips themselves.
A simplified way to see why cluster size and interconnect choice interact:
Cluster compute (illustrative, not an official Huawei formula)
= chip_count × per_chip_throughput × interconnect_efficiency
At 72 chips (Nvidia NVL72 scale):
small chip_count, but near-100% interconnect_efficiency
via tightly coupled electrical NVLink domain
At 15,488 chips (Huawei Atlas SuperPoD scale):
large chip_count, efficiency depends heavily on
the optical fabric holding utilization above the
point where added chips stop adding usable computeThat’s why raw exaflop figures from two very differently sized systems don’t translate into a clean apples-to-apples comparison. A cluster 200 times larger only wins if its interconnect keeps utilization high across every added chip, and that’s precisely the variable Huawei hasn’t published detailed numbers on.
The Real Bottleneck: Manufacturing, Not Demand
Xu’s comment reframes the entire story. This isn’t Huawei being cautious about entering new markets. It’s Huawei admitting it cannot produce enough chips to clear its own backlog. That points squarely at China’s semiconductor manufacturing base, and specifically at SMIC, the foundry Huawei depends on for advanced logic.
The reporting available doesn’t include a verified, current SMIC yield percentage for Ascend-class dies, so any specific figure circulating online should be treated skeptically. What’s well understood industry-wide is the mechanism: without access to the most advanced extreme ultraviolet lithography tools that Dutch supplier ASML is barred from selling to China, SMIC leans on older-generation deep ultraviolet processes pushed harder than they were designed for. That approach can produce workable chips, but typically at lower yields and higher cost per good die than a leading-edge node would deliver.
Layer on top of that the packaging demands of a dual-die part like the 910C, plus high-bandwidth memory supply, plus the optical components the Atlas fabric depends on, and it’s easy to see how “we have enough dies” doesn’t guarantee “we have enough finished, shippable systems.” Every one of those steps is a place where a completed Atlas cluster can stall even when wafer starts look healthy on paper.
How U.S. Export Controls Set the Stage
None of this happened in a vacuum. Washington’s export restrictions on advanced AI chips to China, tightening steadily since 2022, are the backdrop that made Huawei’s Ascend line commercially necessary in the first place. The timeline since 2025 has been unusually volatile even by that standard.
- March 2025: U.S. authorities added Chinese entities tied to advanced AI, supercomputing, and military-linked high-performance computing to the Entity List.
- April 9, 2025: Nvidia disclosed that its China-tailored H20 accelerator now required a U.S. export license, a change that produced a reported $4.5 billion charge tied to H20 inventory and purchase commitments.
- June 2025: Reporting described additional tightening around H20 sales to Chinese buyers.
- July 2025: Policy partially reversed. H20 and AMD’s MI308 were permitted for export under licensing arrangements, with one report citing a 15% revenue-sharing condition attached to the approvals.
- August 2025: Reuters reported licenses being issued for Nvidia’s China exports, alongside disclosure that H20 generated $4.6 billion in first-quarter sales and that China represented roughly 12.5% of Nvidia’s total revenue that quarter.
- 2026: Secondary reporting describes licensing scrutiny extending toward newer Blackwell and Rubin-generation products, though the underlying rules for that extension aren’t fully documented in primary government sources available at publication time.
That back-and-forth is the single biggest reason Chinese buyers hedge toward domestic silicon even when Nvidia hardware is technically available. A supplier that might get cut off by the next policy shift is a supplier you build a second option around, and Huawei has spent three years building itself into exactly that option. It’s the same dynamic playing out on the other side of the ledger in Nvidia’s own Vera Rubin production ramp, where the company is racing to lock in customers before the next round of restrictions can bite.
China’s Domestic AI Chip Push, in Numbers
Beijing has backed the shift toward domestic silicon with procurement preferences, subsidized financing, and pressure on state-linked enterprises to favor Chinese-made accelerators over imported ones. The market-share data reflects how far that push has already moved the needle, though the available figures come from different time periods and measurement methods, so they shouldn’t be read as one continuous trend line.
| Period | Reported Nvidia Share (China) | Reported Domestic Chip Share | Source / Caveat |
|---|---|---|---|
| 2022 | ~95% | Not separately quantified | Cited in Reuters coverage of the market shift; illustrates Nvidia’s earlier dominance |
| 2025 (full-year, AI-accelerator servers) | Not directly given | ~41% | IDC data cited by Reuters; measures server-level accelerator share, not GPU units alone |
| H1 2025 (separate estimate) | ~62% | ~35% | Industry estimate published February 2026; different methodology from the IDC figure above |
| 2026 (unverified industry estimate) | ~8% | Huawei alone estimated at 50-60% of China’s AI chip market | Venture-capital/industry analysis source from July 2026; not a confirmed Huawei or Nvidia financial disclosure |
The direction is unambiguous even where the exact numbers aren’t. Nvidia’s share of its own former stronghold has fallen sharply since 2022, and Chinese accelerator makers, Huawei chief among them, have captured a large and growing share of a market Nvidia essentially owned four years ago. The 2026 figures deserve the most skepticism: a $12.1 billion revenue estimate for Huawei’s AI chip sales and an 8% Nvidia share both trace back to industry-analysis sources rather than audited financial filings, and Huawei doesn’t publish a standalone semiconductor-division revenue line the way a dedicated chip company would.
Competitive Comparison: Huawei, Nvidia, and AMD
Stacking the three companies side by side shows how differently each is positioned right now. Huawei has a captive, subsidized home market and no meaningful legal path to sell its most advanced accelerators broadly overseas, but it also has a genuine manufacturing ceiling limiting how much of that captive demand it can actually serve. Nvidia has the software ecosystem, the manufacturing scale through TSMC, and the efficiency lead Tom’s Hardware pointed to, but its China revenue now moves at the mercy of licensing decisions in Washington. AMD sits in a similar position to Nvidia on export exposure, with its MI308 line reportedly included in the same 2025 licensing changes that reopened limited China sales.
None of the three has a clean advantage. Huawei wins on political certainty inside China and loses on production capacity and efficiency. Nvidia and AMD win on efficiency, software maturity, and manufacturing access, and both remain exposed to a policy environment that has changed direction at least four times since early 2025. That instability is arguably doing more to help Huawei than anything Huawei’s engineers have shipped, because it keeps Chinese customers hedging toward a domestic supplier even when the imported alternative is, chip-for-chip, the better product. It’s the mirror image of what’s happening in the broader Arm-versus-x86 server market, where architecture choice increasingly tracks who controls the supply chain, not just who has the faster part.
Market Impact: Who Feels This First
The most immediate effect lands on Chinese cloud providers and enterprises that had been counting on broader Ascend availability to scale out training and inference capacity this year. If Huawei genuinely can’t clear domestic demand, some of that compute build-out slips, full stop, regardless of budget or willingness to pay.
For Nvidia and AMD, the news is quietly reassuring in the short term. A Huawei that’s capacity-constrained at home is a Huawei that isn’t flooding international markets with cut-rate Ascend hardware, which removes one competitive pressure point outside China even as both companies keep fighting for share inside it. That’s part of why both chipmakers have leaned harder into markets where Chinese competition isn’t yet a factor, including the kind of large sovereign-AI commitments detailed in India’s $12 billion Nvidia Rubin GPU deal with Yotta and hyperscaler-scale orders like AMD’s 50,000-GPU Helios shipment to Oracle.
There’s a longer-term risk buried in the same story, though. Every quarter Huawei spends building and refining Atlas at massive domestic scale is a quarter of real-world engineering experience Nvidia and AMD don’t get a look at. If and when Huawei does eventually clear its capacity backlog, it won’t be launching an unproven first-generation product into export markets. It will be shipping hardware already hardened by running the largest captive AI compute market on earth. Pricing pressure elsewhere in the stack, like the kind visible in Google’s Ironwood TPU pricing against Nvidia, shows how quickly credible alternatives can start reshaping what customers expect to pay once they exist at scale.
Historical Context: China Has Done This Before
The pattern isn’t new. China’s telecom equipment industry followed a similar arc in the 2000s and 2010s: state-favored domestic suppliers, initially behind on core technology, closed the gap over roughly a decade by combining protected home-market volume with sustained state investment, eventually becoming globally competitive exporters. Huawei itself is the standard example from that earlier cycle.
Semiconductors are a harder problem than telecom gear, because chip manufacturing depends on a small number of choke points, extreme ultraviolet lithography chief among them, that China still can’t replicate domestically. That’s the ceiling Xu was effectively acknowledging on September 17. But the demand-side half of the playbook, protected volume plus state backing, is executing on a similar timeline to what worked before. Whether the supply-side half, actual leading-edge manufacturing capability, can close the gap the same way is the open question that will determine whether this month’s export pause is temporary or structural.
Expert Perspective
Huawei’s own leadership has been the most direct source on the company’s reasoning. Eric Xu, Huawei’s rotating chairman, told the audience at Huawei Connect in Shanghai on September 17: “Since we don’t have enough capacity to even satisfy the demand in China, we don’t have a plan to expand into the international market in a fully-fledged way.” Coming from the company’s own rotating chairman rather than a marketing statement, that line functions as a rare, unusually candid admission from a firm that typically frames its AI ambitions in terms of momentum rather than constraint.
What Comes Next: Five Predictions
- Domestic-first stays the policy through at least 2027. Packaging and advanced-node capacity don’t get fixed in a single quarter, so expect Huawei to keep prioritizing Chinese customers well beyond this announcement.
- U.S. export policy keeps swinging. Given four documented direction changes since early 2025, another shift affecting H20, Blackwell, or Rubin licensing to China before mid-2027 looks more likely than a stable, unchanging rulebook.
- Nvidia and AMD chase growth outside China harder. Large sovereign and hyperscaler deals like the ones in India and with Oracle become more central to both companies’ growth stories as China upside stays capped by policy uncertainty.
- Optical interconnect becomes a bigger competitive battleground. As cluster sizes climb toward Huawei’s 15,488-chip territory, the networking fabric linking chips together will draw as much engineering attention as the chips themselves.
- Pressure grows for audited numbers. As unverified estimates about Huawei’s chip revenue and market share keep circulating, expect louder calls from analysts and investors for a real, disclosed breakdown rather than third-party guesses.
What This Means for Buyers and Builders Outside China
If you’re not buying compute in China, the direct effect of this announcement is limited. Huawei’s Ascend hardware was never a legal, practical option for most Western buyers anyway. The indirect effect is worth tracking regardless: a capacity-constrained Huawei is one less source of global GPU price competition, which keeps the pricing floor for Nvidia and AMD hardware roughly where it already sits. Anyone budgeting for AI infrastructure over the next 18 months should treat that as one more reason not to expect a supply-driven price break from outside the existing Nvidia-AMD duopoly any time soon.
Frequently Asked Questions
What did Huawei actually announce about AI chip exports?
Huawei’s rotating chairman Eric Xu said on September 17, 2026, that the company can’t currently meet its own domestic demand for AI computing hardware, so it has no plan to launch a broad international rollout of its Ascend AI chips right now.
What is Huawei’s Atlas SuperPoD?
It’s Huawei’s largest reported AI computing cluster, built from 15,488 Ascend 960 NPUs linked together with optical networking. Huawei claims the configuration reaches up to 30 FP8 exaflops and 120 FP4 exaflops.
How many chips are in Huawei’s largest AI cluster?
According to Tom’s Hardware, the Atlas SuperPoD scales to 15,488 Ascend 960 NPUs in a single system, more than 200 times the chip count of Nvidia’s 72-GPU NVL72 rack.
Why is Huawei limiting international sales of its Ascend AI chips?
Per Eric Xu’s own comments, the constraint is manufacturing capacity, not lack of interest from customers. Huawei’s supply chain, including SMIC fabrication and advanced packaging for dual-die chips like the Ascend 910C, can’t currently produce enough finished systems to clear domestic demand.
How does Huawei’s Atlas cluster compare to Nvidia’s NVL72?
The Atlas SuperPoD has far more chips (15,488 versus 72), but Tom’s Hardware reported that Huawei’s performance-per-watt is expected to trail Nvidia’s by a wide margin, meaning raw chip count and headline exaflop figures don’t translate directly into equivalent real-world efficiency or cost per unit of useful compute.
Is Nvidia still allowed to sell AI chips in China?
Yes, under licensing arrangements that have shifted repeatedly since 2025. The H20 required an export license starting April 2025, was restricted further in June 2025, then partially reopened in July 2025 alongside AMD’s MI308 under a reported revenue-sharing condition. By August 2025, Reuters reported H20 sales at $4.6 billion for one quarter, roughly 12.5% of Nvidia’s total revenue that period.
What is optical interconnect technology and why does it matter for AI chips?
Optical interconnects use light rather than electrical signals to move data between chips, avoiding the bandwidth and distance limits that electrical links hit at large scale. Huawei is using optical networking to hold its 15,488-chip Atlas cluster together, since electrical links alone would struggle to keep that many nodes communicating efficiently.
Will Huawei ever sell its Ascend AI chips outside China broadly?
Possibly, once manufacturing capacity catches up with domestic demand, but Huawei hasn’t given a timeline. Legal export restrictions in most Western markets would also remain a separate obstacle even if Huawei had chips to spare.
