Why Leading AI Firms Are Pursuing Custom Chips in 2026

On July 7, two pieces of news regarding Chinese AI companies' self-developed chips became hot topics.

First, a Reuters exclusive: DeepSeek is developing its own chip. Citing three informed sources, the report states that the chip is targeted at inference, not training. It notes that DeepSeek has been quietly recruiting chip design engineers through non-public channels and has been engaging with external foundries and memory manufacturers. DeepSeek has not responded to requests for comment, and the report has sparked some skepticism, as it currently relies solely on Reuters as a source.

On the same day, The Information reported that another Chinese lab, Zhipu, is also evaluating its own custom-designed chips. According to sources familiar with the matter, the reason is the surging demand for GLM-5.2. As reported by The Information and multiple outlets citing it, this model has become the fastest-growing on the Vercel model aggregation platform, with daily token consumption peaking at 27 times its initial level during its first week of launch. Zhipu has initiated preliminary discussions with several domestic chip design companies but has not yet selected a partner; the project is reportedly expected to take more than two years to complete. Following the news, Zhipu’s stock in Hong Kong rose as much as 9.9% on the day.

Of course, both messages are still in the early evaluation stage, with no physical products or finalized designs.

In fact, DeepSeek and Zhipu are not alone. Zoom out, and you’ll see that “AI companies building their own chips” has, by 2026, shifted from an occasional choice to an industry-wide norm: from across the ocean at OpenAI and Anthropic, to these two leading labs domestically, the actions are remarkably aligned. This article aims to answer why.

If news about DeepSeek remains at the rumor stage, OpenAI has already put the real thing on display.

On June 24, OpenAI and Broadcom jointly released Jalapeño, OpenAI’s first in-house chip, also an ASIC (application-specific integrated circuit) designed specifically for large model inference.

Richard Ho, Head of Hardware at OpenAI, said the chip was "designed from the ground up for LLM inference," with targeted optimizations around the core, memory movement, networking, and service patterns.

According to OpenAI, early lab tests show that Jalapeño's performance per watt "significantly surpasses the current industry leading level." However, the company also acknowledges that final performance is still being measured, and detailed technical reports will not be released until future months. Until third-party benchmarks are available, such performance claims by manufacturers remain unverified.

Several details are worth noting. According to Tom's Hardware, the chip went from design to tape-out in just nine months, and OpenAI claims this may be the fastest ASIC development cycle in the history of high-performance advanced semiconductors; accelerating this cycle was OpenAI’s own model: using AI to design chips that build AI.

Jalapeño is scheduled to begin deployment by the end of 2026, anchored by a Broadcom partnership involving a 10 GW capacity set to be completed by 2029. Microsoft is reportedly expected to purchase approximately 40% of Jalapeño’s initial capacity.

Now consider another cutting-edge lab, Anthropic. In April this year, Reuters was the first to report that Anthropic was also evaluating the development of its own chips, using cautious language: the plan is in an early stage, and the company might ultimately decide to purchase rather than build, with no finalized design or dedicated team assembled yet.

But by early July, new developments emerged. According to multiple media reports, Anthropic has begun engaging with Samsung to explore manufacturing a custom chip, reportedly targeting Samsung’s 2-nanometer process and advanced packaging.

And Anthropic recently recruited Clive Chan, an early member of OpenAI’s in-house chip team.

In response to further inquiries, Anthropic stated that its computing strategy will continue to center on a diversified hardware stack composed of chips from Google, Amazon, and NVIDIA, and declined to provide further comment on its partnership with Samsung.

This official statement precisely highlights the most authentic motivation behind the wave of in-house chip development. According to data from The Information, NVIDIA controls approximately 74% of the global AI chip market; Anthropic has never manufactured its own chip, and every invocation of Claude runs on chips rented from partners—partners who are also its competitors.

Notably, Anthropic’s exploration of developing its own chips coincided almost exactly with a sharp increase in its revenue. According to its own disclosures, annualized operating revenue had surpassed $30 billion by 2026, up from approximately $9 billion at the end of 2025. At this scale, the economics of developing custom chips finally make sense.

When you compare the announcements from DeepSeek, Zhipu, OpenAI, and Anthropic side by side, you’ll notice a common thread: they’re all building inference chips, not training chips. This is no coincidence.

A structural shift currently underway in the industry is that the focus of computational power consumption is moving from "training models" to "running models." Training is a one-time cost, whereas serving models to hundreds of millions of users incurs ongoing expenses. According to Introl’s industry analysis, inference now accounts for approximately two-thirds of all AI compute power.

https://introl.com/blog/custom-silicon-inflection-2026-hyperscaler-asics-nvidia-gpu

And inference is precisely ASICs’ domain. As Chris Miller, author of The Chip War, told CNBC: NVIDIA’s GPUs are like a Swiss Army knife—capable of handling any parallel computing task—while ASICs are like single-purpose tools: highly efficient and fast, but hardwired to perform only one type of job. Training requires the flexibility of a Swiss Army knife, since model architectures are still evolving; but once the model is finalized and needs to serve massive volumes of requests, that single-purpose tool becomes more energy-efficient and cost-effective.

https://oplexa.com/custom-asic-market-2026-hyperscalers-ditching-nvidia/

The surge in agents has further amplified this ledger. An analysis by MindCast AI presents an interesting distinction:

Traditional reasoning involves a "query cost": one question, one answer, settled.

Agents are "cyclical in cost": a single user goal can trigger dozens or even hundreds of inference calls, as the agent must reason, plan, retrieve, and execute.

Image translated from https://www.mindcast-ai.com/p/ai-inference-economy

As agents are deployed at scale, the economic asymmetry between training and inference will grow nonlinearly. Therefore, this analysis firm concludes that inference economics is not a concern for 2028—it’s a procurement decision that must be made in 2026.

For "why build it yourself," the industry already has quite specific data to support it.

One of the most straightforward examples comes from the AI image platform Midjourney. According to reports, after migrating its inference workload from NVIDIA GPUs to Google’s seventh-generation TPUs, Midjourney reduced its monthly compute costs from approximately $2.1 million to about $700,000—a 65% decrease. When scaled up to hyperscale operators running billions of queries daily, as one analyst noted, investing billions of dollars in developing proprietary chips becomes a straightforward financial calculation.

The answer to this question is transforming the entire market. According to TrendForce’s forecast, the shipment growth rate of custom ASICs will reach 44.6% by 2026, compared to just 16.1% for commercial GPUs—marking the first time custom chip growth surpasses that of GPUs. Introl cites even broader figures from Bloomberg Intelligence: by 2033, the entire AI accelerator market is projected to reach $604 billion, with the share of custom silicon continuing to accelerate.

Large-scale manufacturers have already voted with real money. Google’s TPU, Amazon’s Trainium, Microsoft’s Maia, and Meta’s MTIA—each of these custom chips serves massive inference workloads within their parent companies.

https://oplexa.com/custom-asic-market-2026-hyperscalers-ditching-nvidia/

According to CNBC, Ron Diamant, Chief Architect of Trainium, said that Amazon’s ASIC offers a 30% to 40% cost-performance advantage over other hardware providers on AWS.

These figures show that when your computing power demands scale to hypersize, chips are no longer a cost item—they become the very competitive barrier. As an Oplexa analysis states, controlling your own silicon means controlling your own performance roadmap, cost structure, and supply chain—three things no purchase order can buy.

If the economic argument is the surface-level reason, then "not wanting to entrust one's fate to a single company" is the deeper current beneath the wave.

NVIDIA’s strength stems partly from the chips themselves, but even more from CUDA—a software ecosystem built up over more than two decades. According to Spheron’s analysis, nearly all serious LLM inference optimizations—from FlashAttention to continuous batching in vLLM—run exclusively on CUDA. This ecosystem is less a moat for NVIDIA and more a “migration cost” that any team looking to switch chips must pay. Estimates suggest that porting a service stack from vLLM to Amazon’s Neuron SDK often takes two to six weeks of engineering effort, and some model architectures aren’t supported at all.

https://www.spheron.network/blog/hyperscaler-custom-ai-chips-2026-trainium-tpu-maia-mtia-vs-nvidia-gpu/

Precisely because this wall is so tall and thick, the motivation to bypass it is especially strong. For hyperscale manufacturers and cutting-edge labs, every workload running on their own custom chips means more profit retained in their own pockets and greater leverage when negotiating with NVIDIA.

TheStreet’s assessment is straightforward: what’s changing is the default assumption that AI labs must accept NVIDIA’s any price and supply terms. Microsoft has Maia, Amazon has Trainium, Google has TPU, and the latest wave of AI labs now wants to build their own similar leverage.

Return to the two messages from July 7. For DeepSeek and Zhipu, in addition to all the reasons mentioned above, they jointly face an issue that their U.S. counterparts do not: export controls.

It is also worth noting that the rumors surrounding DeepSeek’s in-house chip development coincide precisely with its first external funding round: the company raised over 50 billion yuan, achieving a valuation exceeding 330 billion yuan. Chip development is extremely capital-intensive; the timing of these two events—funding and chip development—is unlikely to be coincidental.

Zhipu, meanwhile, became the first Chinese AI lab to list on the Hong Kong Stock Exchange back in January this year. Both are preparing their finances in advance for this expensive gamble.

Pulling these clues together, the answer to “Why do AI companies all want to build their own chips?” is the convergence of several forces: inference has replaced training as the main battleground for computing power, creating a role for specialized ASICs; the explosion of agents has pushed inference costs into nonlinear growth, magnifying the value of every dollar saved; and deep reliance on NVIDIA has driven everyone to build their own bargaining power. For Chinese companies like DeepSeek and Zhipu, export controls add another layer to this equation.

But beneath the hype, a reality check is needed. Developing in-house chips is far from a guaranteed win. Design cycles can take 18 to 24 months, require massive upfront engineering investment, and only make sense if workloads are sufficiently stable and predictable. For startups still experimenting with model architectures or ordinary businesses with diverse workloads, NVIDIA GPUs’ flexibility remains the more cost-effective choice.

TechTimes clearly states that the 40% to 65% cost advantage applies only to giants conducting billions of queries per day; for companies running only tens of thousands of queries per week, the math is reversed.

A more compelling question to ask is: Will this wave weaken NVIDIA, or merely add a complementary role for it? Most analyses lean toward the latter. According to Oplexa, ASIC shipments may surpass GPU shipments in volume by 2027, but both markets will continue to grow: AI infrastructure is splitting into two distinct paths—ASICs for fixed, high-frequency, predictable workloads, and GPUs for research-intensive, diverse, and still-evolving architectures. NVIDIA still firmly controls training and the vast majority of the market.

https://oplexa.com/custom-asic-market-2026-hyperscalers-ditching-nvidia/

It is still unclear whether DeepSeek and Zhipu can develop their own chips, as both reports remain in the early evaluation stage.

But they landed on the same day, like two stones cast simultaneously into water, their ripples revealing the same anxiety: from San Francisco to Beijing, in this era where computational power equals authority, no AI company hoping to survive to the next round is willing to lease its lifeline indefinitely in someone else’s hands.

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by Panda.