Huawei just moved up the clock on its next flagship AI chip, and the timing lines up with a separate story that shows why the company is in a hurry. According to TechCrunch, Huawei announced at its Huawei Connect conference on September 17, 2026 that it is pulling forward the launch of its next-generation Ascend 960DT chip to the first quarter of 2027, months earlier than the third-quarter 2027 window it had previously set. Days earlier, Moneycontrol reported that Chinese AI lab DeepSeek is planning to deploy at least 160,000 Huawei accelerators at a data center under construction in Inner Mongolia, a cluster that could become one of the largest known concentrations of Huawei AI chips anywhere.
Put those two facts together and the picture is clear in this fast-moving corner of the AI chips market: China’s largest AI players are betting real infrastructure money on Huawei silicon, and Huawei is racing to ship a chip good enough to justify that bet before Nvidia’s next platform ships in volume. This is not an isolated China story either. The same week brought reports that Microsoft, Meta, OpenAI, and Amazon are all pushing their own custom AI chips into production, a signal that the entire industry is trying to loosen its dependence on Nvidia at the same time.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Huawei Moves Its Next AI Chip Launch Up by Two Quarters
The headline move came from Huawei Connect, the company’s annual flagship event, where executives detailed a revised roadmap for the Ascend line in September 2026. TechCrunch reported that the Ascend 960DT, originally slated for a third-quarter 2027 release, is now targeted for the first quarter of 2027, a pull-forward Huawei itself has characterized as three full quarters earlier than its original internal schedule. A Huawei spokesperson framed the change as part of the company’s broader push to challenge Nvidia in AI accelerators, according to the outlet’s coverage of the event.
Separately, Tom’s Hardware noted that Huawei used the same roadmap update to detail performance targets further out on the line, including the Ascend 960PR, whose FP4 (4-bit floating point) throughput is now expected to double prior projections. Huawei has also said the broader 960 generation will roughly double compute and memory capacity compared with the current Ascend 950 class, a jump the company attributes to the same September 2026 architectural changes. FP4 matters because low-precision math is what most large-scale AI inference and a growing share of training now runs on, so a jump there translates fairly directly into more tokens processed per chip per dollar spent on power and rack space.
None of this makes the Ascend 960DT a confirmed Nvidia-beater. Huawei has not published independently verified benchmark comparisons against Nvidia’s current data-center lineup, and the revised launch date is still roughly a year and a half away. What has changed is the urgency behind the schedule, and that urgency traces directly back to demand already lined up in China.
Inside DeepSeek’s 160,000-Chip Order for Inner Mongolia
DeepSeek, the Chinese AI lab that rattled global markets with its efficient open-weight models, is reportedly planning to deploy at least 160,000 Huawei AI accelerators at a new data center it is building in Inner Mongolia. Moneycontrol’s report describes the buildout as part of a broader Chinese push to reduce reliance on Nvidia, driven in large part by years of tightening U.S. export controls that have restricted the sale of Nvidia’s most advanced data-center GPUs into China.
A cluster of that size would put DeepSeek among the largest known domestic deployers of Huawei’s Ascend chips, alongside state-linked cloud operators and other major Chinese AI labs. It also tells a specific story about confidence: DeepSeek built its reputation on squeezing more performance out of constrained hardware, most visibly with the V3 and R1 model families that showed strong results despite running on chips well below Nvidia’s top tier. That confidence looks less like a stretch once you account for Huawei’s own scaling ambitions: the company said in September 2026 that its Peerium interconnect technology is designed to support clusters of up to 1 million processors, well beyond the 160,000-chip scale DeepSeek is reportedly targeting. A large-scale bet on Huawei silicon suggests DeepSeek’s engineers believe they can get comparable training and inference throughput out of the Ascend platform, provided the software stack cooperates.
That last condition is the one to watch. Chip announcements are easy. Getting a large, heterogeneous cluster of 160,000 accelerators to train frontier-scale models reliably, at high utilization, is a systems engineering problem that has taken Nvidia and its CUDA ecosystem more than a decade to refine.
The Ascend Roadmap: From 950DT to 960PR to 960DT
Huawei’s Ascend line has moved fast since U.S. export restrictions cut off most of its access to Nvidia and other Western AI silicon. The current top-end chip in wide deployment is the Ascend 950DT class, which according to a September 11, 2026 Asia Intelligence Brief report from Rio Times has seen its price pass 250,000 yuan per unit on the Chinese market, roughly the price range analysts associate with high-end data-center accelerators once memory and packaging costs are included.
The same report noted that a shortage of high-bandwidth memory has lifted Chinese AI chip prices broadly by 20 to 50 percent, that Nvidia itself has flagged server price increases above 15 percent, and that contract manufacturer Foxconn posted a record August on the back of AI hardware demand. In other words, Huawei is trying to accelerate a chip launch at the exact moment its own input costs, memory chief among them, are climbing sharply.
What the 960DT and 960PR Actually Change
Tom’s Hardware’s coverage of the roadmap update described the Ascend 960PR’s FP4 performance target as double what was previously expected, a jump Huawei attributes to architectural changes disclosed at Huawei Connect. The 960 generation is now split into two dedicated parts: the Ascend 960DT, getting the accelerated Q1 2027 launch, is positioned as the inference-focused chip, while the Ascend 960PR, slotted for Q3 2027, is the dedicated training part. Huawei has since disclosed concrete specifications for the 960DT, including 2 FP8 PFLOPS and 4 FP4 PFLOPS of compute, 288 GB of onboard memory, 9.6 TB/s of memory bandwidth, and 2.2 TB/s of interconnect bandwidth, figures that move the chip from vague roadmap language to an actual spec sheet, even though independent, third-party benchmark testing to validate those numbers has not yet happened.
Why Nvidia Is Effectively Locked Out of This Race
None of this would matter as much if Nvidia’s newest data-center GPUs were freely available in China. They are not. Multi-year U.S. export controls have progressively restricted the most capable Nvidia accelerators from being sold into the Chinese market, which is precisely the vacuum Huawei, DeepSeek, and other domestic players are moving to fill. Huawei says that vacuum is already being filled at real scale: as of September 2026, the company reports having shipped more than 1,000 AI supernodes to over 370 customers. That dynamic is why a 160,000-chip order for Huawei silicon reads as strategic rather than purely a cost decision. For a lab the size of DeepSeek, Nvidia’s top-tier hardware is largely off the table as a large-scale buildout option inside China, leaving Huawei’s Ascend line as the most credible domestic alternative at scale.
That reality cuts both ways. It guarantees Huawei a captive market of sorts, but it also means the company has to prove its chips can handle frontier-scale training workloads without the fifteen-plus years of software tooling Nvidia has built around CUDA. The pace of Huawei’s roadmap updates, three chip generations detailed at a single event, suggests the company understands the software gap is the harder problem and is trying to out-iterate it on hardware while its ecosystem catches up.
A Five-Front Custom Silicon War: Microsoft, Meta, OpenAI, Amazon and Huawei
Huawei’s roadmap update did not happen in a vacuum. The same week, multiple outlets reported that the biggest U.S. cloud and AI companies are simultaneously pushing their own custom chips into production, each trying to cut how much of their infrastructure spend flows to Nvidia.
According to a September 7, 2026 report from Sedaily, Microsoft could unveil its Maia 300 AI accelerator as early as this month and is negotiating to secure more than 300,000 units for delivery next year, positioning Maia as a large-scale in-house alternative to Nvidia GPUs for Azure’s AI services. Sedaily’s reporting also described Meta’s Iris chip, OpenAI’s Jalapeño processor, and an Amazon-Qualcomm partnership as part of the same coordinated shift toward custom or co-developed AI hardware.
Separate coverage from industry publication OpenPR detailed each of those efforts. Meta’s Iris chip, built under its MTIA (Meta Training and Inference Accelerator) program, is scheduled to enter production this month. OpenAI’s Jalapeño processor, developed with Broadcom and manufactured by TSMC, has moved into a broader production strategy. And Qualcomm and Amazon announced a multigenerational agreement in September 2026 to supply customized AI chips for Amazon’s data centers, extending a deal already reported at tens of billions of dollars in scale.
Sedaily’s analysis was blunt about the motivation: software giants are betting on custom AI chips despite a real hardware performance gap against Nvidia, because the alternative, staying fully dependent on one supplier during a period of explosive AI infrastructure demand, is judged to be the riskier position.
What Ties These Five Efforts Together
Strip away the geography and the five programs share a common logic. Each company is the single biggest buyer of its own future chip, which removes the margin a merchant supplier like Nvidia needs to build in. Each is targeting internal workloads first rather than trying to sell chips externally, which lowers the bar for a first-generation part to be useful even if it trails Nvidia on raw performance. And each announcement landed within roughly two weeks of the others in September 2026, which suggests the underlying driver, tight and pricier Nvidia supply amid surging AI demand, is being felt industry-wide at the same time rather than by one company in isolation.
HBM Shortage Is Squeezing Everyone, Including Huawei
High-bandwidth memory, the stacked DRAM that feeds data into AI accelerators fast enough to keep them from sitting idle, is in short supply across the entire industry. Rio Times reported Chinese AI chip prices climbing 20 to 50 percent as a direct result of the HBM shortage, with Nvidia itself warning of server price increases above 15 percent. That shortage does not discriminate by nationality of the chip design. Huawei’s Ascend chips need HBM just as much as Nvidia’s do, and Huawei has less direct access to the leading memory suppliers, SK Hynix, Samsung, and Micron, than Nvidia does.
That constraint is part of why the pricing on the current Ascend 950DT crossing 250,000 yuan matters. It shows Huawei’s chips are not immune to the same memory-driven cost pressure squeezing every AI hardware maker right now, even as the company accelerates its next-generation launch. A tighter memory market could offset some of the benefit of shipping the 960DT two quarters earlier if the chip can’t be produced in volume at a competitive price.
| Metric | Reported Figure | Source |
|---|---|---|
| Chinese AI chip price increase (HBM-driven) | 20 to 50 percent | Rio Times, Asia Intelligence Brief, Sept. 11, 2026 |
| Ascend 950DT unit price | Passed 250,000 yuan | Rio Times, Asia Intelligence Brief, Sept. 11, 2026 |
| Nvidia flagged server price increase | Above 15 percent | Rio Times, Asia Intelligence Brief, Sept. 11, 2026 |
| Foxconn monthly revenue | Record August, driven by AI hardware demand | Rio Times, Asia Intelligence Brief, Sept. 11, 2026 |
| DeepSeek Huawei chip order (Inner Mongolia) | At least 160,000 accelerators | Moneycontrol |
| Microsoft Maia 300 unit target (2027) | More than 300,000 units in negotiation | Sedaily, Sept. 7, 2026 |
Nvidia’s Countermove: Rubin, NVLink Fusion and Third-Party Chips
Nvidia is not standing still while its customers and rivals build alternatives, and it still controls the vast majority of global GPU shipments. The company’s Rubin platform, introduced earlier in 2026 as the successor to its Blackwell-era lineup, pairs a Rubin GPU with a Vera CPU, NVLink 6 switching, ConnectX-9 networking, and BlueField-4 data processing units into a full rack-scale system rather than a standalone chip.
More notably, Nvidia is opening parts of that platform to outside chipmakers. According to Semiconductor Engineering‘s chip industry roundup, MediaTek will offer Nvidia’s NVLink Fusion interconnect as a design foundation for custom AI accelerator customers, letting third-party chipmakers plug their own silicon into Nvidia’s MGX rack architecture rather than being frozen out of the ecosystem entirely. That is a defensive move as much as an offensive one: if hyperscalers are going to build custom chips regardless, Nvidia would rather those chips talk to its networking and rack infrastructure than to a rival’s.
Market Impact: What This Means for Cloud Pricing and GPU Supply
For buyers of cloud AI compute, the immediate effect of this custom-silicon wave is more supply diversity, at least on paper. If Microsoft can bring on hundreds of thousands of Maia accelerators, and Meta and Amazon ramp their own chips at the same time Huawei accelerates Ascend, the theoretical pool of AI compute grows faster than Nvidia alone could ship it. In practice, most of that new capacity is going toward each company’s internal workloads first, Meta’s own model training, Microsoft’s Azure AI services, Amazon’s internal and Anthropic-linked demand, rather than showing up as cheaper general-purpose GPU rental overnight.
The HBM shortage is arguably the bigger near-term price driver for buyers outside China. With Nvidia flagging server price increases above 15 percent and Chinese chip prices up 20 to 50 percent on the same memory constraint, anyone shopping for AI training or inference capacity in late 2026 is contending with a tighter, pricier market regardless of which chip vendor they choose. Diversifying the supplier list helps long-term negotiating leverage, but it does not fix a memory bottleneck that sits upstream of every chipmaker on this list.
Historical Context: How Huawei Got Here After Export Controls
Huawei’s AI chip program exists largely because of policy, not just ambition. Since the U.S. began restricting exports of advanced semiconductors and chipmaking tools to Chinese firms, Huawei has had to build an AI accelerator business without reliable access to the most advanced foreign fabrication processes or design tools that Nvidia, AMD, and other Western chipmakers rely on, even as those same chipmakers pursue their own deeper hardware partnerships elsewhere. The Ascend line’s earlier chips were built under those constraints, and the pace of the current roadmap, moving from the 950DT class already on the market to a 960 generation with a 960PR and 960DT variant detailed in the same announcement, shows a company trying to compress a multi-generation product cycle into a shorter window under pressure from both geopolitics and domestic demand.
That compression carries risk. Moving a chip launch up by two quarters, as Huawei has done with the 960DT, usually means less time for validation, yield tuning, and software optimization before mass production. Whether Huawei can hit the new Q1 2027 date without compromising reliability at scale is an open question that will not be answered until real deployments, including whatever share of DeepSeek’s 160,000-chip Inner Mongolia cluster runs on 960-series silicon, are up and running.
Huawei Ascend vs Nvidia vs US Hyperscaler Chips: A Side-by-Side Look
The table below lines up the major custom and merchant AI accelerator efforts reported as active in September 2026. Specifications beyond what each company or outlet has disclosed publicly are marked as unconfirmed rather than estimated.
| Chip / Platform | Company | Reported Status (Sept. 2026) | Primary Buyer | Source |
|---|---|---|---|---|
| Ascend 960DT | Huawei | Launch moved up to Q1 2027, from Q3 2027 | Chinese cloud and AI labs, including DeepSeek | TechCrunch |
| Ascend 960PR | Huawei | FP4 performance target doubled vs. prior roadmap | Chinese data centers | Tom’s Hardware |
| Ascend 950DT | Huawei | Already shipping; price passed 250,000 yuan | Domestic Chinese buyers | Rio Times |
| Rubin platform | Nvidia | Introduced in 2026 as Blackwell-era successor | Global hyperscalers | Semiconductor Engineering |
| Maia 300 | Microsoft | Unveiling expected as early as Sept. 2026; 300,000+ units in negotiation for next year | Microsoft Azure | Sedaily |
| Iris | Meta | Entering production Sept. 2026 under MTIA program | Meta internal AI workloads | OpenPR |
| Jalapeño | OpenAI (with Broadcom, built by TSMC) | Moved into broader production strategy | OpenAI internal workloads | OpenPR |
| Custom Qualcomm accelerator | Qualcomm / Amazon | Multigenerational supply agreement announced Sept. 2026 | Amazon data centers | OpenPR |
The Software Moat Problem: CUDA vs CANN
Hardware specs get the headlines, but the harder problem for every company on that table, Huawei included, is software. Nvidia’s CUDA ecosystem has more than fifteen years of accumulated libraries, framework integrations, and developer familiarity behind it. Huawei’s equivalent, its CANN (Compute Architecture for Neural Networks) toolkit, is younger and has a smaller pool of engineers who know it well outside China.
DeepSeek’s willingness to commit 160,000 chips to a Huawei-based cluster is notable precisely because it signals confidence that this software gap is closeable, at least for the kinds of large language model training and inference workloads DeepSeek runs. If DeepSeek can demonstrate strong utilization and training stability on Ascend hardware at that scale, it would be one of the strongest public proof points yet that a non-Nvidia AI stack can run frontier-class workloads reliably. If it struggles with driver stability, kernel optimization, or interconnect bottlenecks at scale, that would reinforce how durable Nvidia’s software advantage remains even as hardware alternatives multiply.
Five Predictions for the Custom AI Chip Race Through 2027
- Huawei will face pressure to hold its new Q1 2027 date for the Ascend 960DT, and any slip will be read as evidence that its accelerated roadmap outran its manufacturing and yield capacity.
- DeepSeek’s Inner Mongolia cluster will become a closely watched proof point for whether large-scale LLM training is viable on non-Nvidia hardware at a scale that matters commercially, not just in benchmarks.
- Microsoft’s Maia 300 rollout, if it reaches the reported 300,000-plus unit range, will make Microsoft one of the largest non-Nvidia AI silicon deployers among U.S. hyperscalers by volume.
- The HBM shortage will remain the dominant near-term price driver for AI compute industry-wide, likely mattering more to buyers’ bills in the next two quarters than any single vendor’s new chip launch.
- Nvidia’s decision to open NVLink Fusion to third-party chips such as MediaTek-designed silicon suggests the company expects the custom-chip trend to keep growing and is positioning to stay the connective layer even where it isn’t the compute vendor.
What This Means for Developers and Cloud Buyers
For software engineers building on top of cloud AI infrastructure outside China, the practical impact of this week’s news is indirect for now. Nvidia GPUs remain the default target for most training and inference frameworks, and that is not changing in the short term. But the direction of travel is worth tracking. Microsoft, Meta, OpenAI, and Amazon are all now running or ramping their own silicon for at least part of their internal workloads, which means the frameworks and compilers teams rely on, PyTorch, JAX, and vendor-specific inference runtimes, will increasingly need to support more than one hardware target well.
Engineers working with Chinese cloud providers or evaluating DeepSeek’s models should watch for any published details on training infrastructure. If DeepSeek discloses meaningful detail about running on Huawei’s Ascend cluster, that would be one of the more concrete public data points yet on how a non-Nvidia stack performs at frontier scale, useful information for anyone assessing hardware risk in their own AI infrastructure planning.
Risks and Open Questions
Several things about this story remain unconfirmed. Huawei has not published independent, third-party-verified benchmarks for the Ascend 960DT or 960PR comparing them directly to Nvidia’s current or upcoming chips. The DeepSeek cluster figure of 160,000 chips comes from Moneycontrol’s reporting and has not been confirmed by DeepSeek or Huawei directly in the material reviewed for this article. And the accelerated Q1 2027 launch date for the 960DT is a company-announced target, not a shipped product, meaning it is subject to the same manufacturing, yield, and supply chain risks that have pushed back chip launches across the industry before.
The HBM shortage adds another layer of uncertainty. If memory supply tightens further before Q1 2027, Huawei could face the same kind of component-driven delay that has already pushed other companies’ hardware roadmaps around in 2026. None of that erases the significance of the announcement, but it does mean the real test of this story is still roughly a year and a half away.
Frequently Asked Questions
What is the Huawei Ascend 960DT?
It is Huawei’s next-generation AI accelerator chip, part of the Ascend 960 family, which Huawei detailed at its Huawei Connect conference on September 17, 2026. It is designed to compete with Nvidia’s data-center AI GPUs.
When is the Ascend 960DT launching?
Huawei moved the launch up to the first quarter of 2027, from a previous target of the third quarter of 2027, according to TechCrunch’s coverage of the Huawei Connect announcement.
How many Huawei chips is DeepSeek reportedly buying?
Moneycontrol reported that DeepSeek plans to deploy at least 160,000 Huawei AI accelerators at a data center it is building in Inner Mongolia.
Why can’t DeepSeek just use Nvidia chips?
Years of U.S. export controls have progressively restricted the sale of Nvidia’s most advanced data-center GPUs into China, pushing Chinese AI labs toward domestic alternatives like Huawei’s Ascend line for large-scale deployments.
Is Huawei’s Ascend chip faster than Nvidia’s chips?
There is no independently verified, direct benchmark comparison available. Huawei has said the Ascend 960PR’s FP4 performance target is double what was previously expected on its own roadmap, but that is a company-disclosed target rather than a third-party test result.
What other companies are building custom AI chips right now?
Microsoft (Maia 300), Meta (Iris), OpenAI (Jalapeño, built with Broadcom and TSMC), and Amazon (with Qualcomm) are all reported to be moving custom AI accelerators into production around the same time as Huawei’s roadmap update, according to Sedaily and OpenPR.
Why are AI chip prices rising in 2026?
A shortage of high-bandwidth memory (HBM) has pushed Chinese AI chip prices up 20 to 50 percent and led Nvidia to flag server price increases above 15 percent, according to Rio Times’ Asia Intelligence Brief.
Does this threaten Nvidia’s dominance in AI chips?
Not immediately. Nvidia still leads on software ecosystem maturity through CUDA and is expanding its Rubin platform and NVLink Fusion interconnect to stay central even as customers diversify hardware. But the simultaneous push from Huawei, Microsoft, Meta, OpenAI, and Amazon shows the industry is actively working to reduce single-vendor dependence.
Related Coverage
- iPhone 18 Pro Max QLC Storage Slower Than SD Card [2026]
- Arctis Nova Pro vs Stealth 600 vs BlackShark V3 Pro: $250 Gap [2026]
- PBT vs ABS vs POM Keycaps: 5-Year Shine Gap Tested [2026]
- ARC X1 Handheld Launches at Rs 29,500, Undercuts Rivals [2026]
- How to Set Up Gyro Aiming: 12 Steps, 45 Min [2026]
