The New AI Chip Race: Model Makers Turn to Custom Silicon

July 2026 marks a turning point in the AI industry. For years, the dominant narrative was simple: AI labs trained models on Nvidia GPUs and paid per token for inference. That model is now breaking apart. In just the past few weeks, DeepSeek, OpenAI, Anthropic, and Meta have all advanced their custom AI chip plans — signaling a fundamental shift from "buying compute" to "defining compute."

OpenAI Jalapeño: First Custom Inference Chip

On June 24, 2026, OpenAI unveiled Jalapeño, its first custom Intelligence Processor — an accelerator designed from the ground up for large language model inference. The chip was co-developed with Broadcom and Celestica in just nine months, a timeline OpenAI describes as potentially the fastest ASIC development cycle ever achieved in high-performance semiconductors [citation:4][citation:9].

Jalapeño is not a general-purpose accelerator. It was designed around OpenAI's deep understanding of LLM fundamentals, informed by its roadmap of models, kernels, serving systems, and product needs. Engineering samples are already running ML workloads in the lab at production target frequency and power, including GPT-5.3-Codex-Spark [citation:4].

Early testing shows the chip will deliver "performance per watt substantially better than current state-of-the-art," according to OpenAI's announcement [citation:4]. The architecture reduces data movement and balances compute, memory, and networking resources to achieve utilization much closer to theoretical peak performance.

The chip is designed for flexibility to work with all LLMs, guided by OpenAI's insights into inference needs across the industry. It is the first step in a multi-generation compute platform, with initial deployment planned by the end of 2026 at gigawatt scale with data center partners including Microsoft [citation:4][citation:9].

Key specs and details:

Greg Brockman, President and Co-Founder of OpenAI, described the strategic importance: "The world is moving to a compute-powered economy. Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses" [citation:4].

DeepSeek Enters the Chip Game

On July 7, 2026, Reuters reported that DeepSeek is developing its own AI inference chip, with the project having started approximately one year ago [citation:1][citation:14]. The chip is designed specifically for inference — the process of generating responses to user queries — not for training.

This move represents a significant strategic pivot for the Chinese AI company. DeepSeek's journey has been defined by navigating chip restrictions: it trained its R1 model on Nvidia H800 chips before they were banned, then adapted its V4 model to Huawei Ascend architecture, and is now designing its own silicon [citation:1].

According to sources cited by Reuters, DeepSeek has engaged with chip design firms, foundries, and memory suppliers. The company has been quietly expanding its chip design team over recent months through headhunting and internal referrals — no public job postings were used [citation:1].

The project is backed by DeepSeek's first-ever external funding round completed in June 2026, raising approximately 51 billion yuan ($7.4 billion) at a valuation between $52 billion and $59 billion. The funding is explicitly allocated toward expanding compute centers (primarily with domestic chips), developing its own AI chips, and recruiting global talent [citation:1].

Key details:

The motivation is clear: DeepSeek's aggressive pricing strategy — V4 Flash at 2 yuan per million output tokens, less than 1/20th of GPT-5.5 — has generated massive usage. But with tens of millions of daily active users, every token costs money. At that scale, even small efficiency gains translate into billions of dollars in savings [citation:1].

Industry analysts view this as a natural progression for China's AI champion. DeepSeek's chip journey mirrors the broader trajectory of China's AI industry: from reliance on Nvidia, to adaptation on Huawei, to ultimately designing its own silicon [citation:1][citation:6].

Anthropic Explores Custom Silicon with Samsung

On July 2, 2026, The Information reported that Anthropic is in early discussions with Samsung Electronics about potential manufacturing partnerships for custom AI chips [citation:2][citation:7][citation:11]. The project remains in very early stages — Anthropic has not yet finalized the chip's design, processing function, performance specifications, or deployment architecture [citation:2].

According to sources, Anthropic is considering Samsung's 2nm process technology and advanced packaging solutions. The company has engaged with multiple chip design firms but has not yet entered detailed design, testing, or manufacturing phases [citation:7][citation:15].

The move follows Anthropic's aggressive hiring: in June 2026, the company brought on Clive Chan, an early member of OpenAI's custom chip team who previously worked on Tesla's Dojo supercomputer project [citation:15].

Key details:

Anthropic's official response emphasizes that AWS Trainium, Google TPU, and Nvidia GPUs will remain central to its compute strategy [citation:11]. The company is not abandoning its multi-supplier approach — rather, it's adding a fourth option. This reflects the unique position Anthropic has carved out: while OpenAI is tied to Nvidia and xAI to its own infrastructure, Anthropic has been hedging across multiple hardware platforms [citation:15].

But the economics are pushing toward custom silicon. Anthropic's annualized revenue run rate has grown from $9 billion at the end of 2025 to over $47 billion by May 2026 — a five-fold increase in five months [citation:15]. With that kind of scale, even a few percentage points of efficiency translate into billions in savings. And as the company acknowledged, rapid growth creates "inevitable pressure" on infrastructure [citation:15].

Meta Iris: Production in September

On July 9, 2026, Reuters reported that Meta's custom AI chip, code-named Iris, will enter production in September 2026 [citation:3][citation:12]. An internal memo reviewed by Reuters shows the chip cleared six weeks of testing without any major issues — a notable milestone for a program that has struggled for years [citation:12].

Iris is part of Meta's Training and Inference Accelerator (MTIA) program, which includes four planned chip generations: MTIA 300, 400, 450, and 500. The company plans to release a new generation approximately every six months through 2027, roughly twice the speed of a typical chip development cycle [citation:3][citation:12].

Broadcom is the design partner on Iris, and Taiwan Semiconductor Manufacturing Co. (TSMC) is the foundry. Meta has also signed long-term supply agreements with Samsung Electronics for memory chips, Sandisk for flash storage, and Sumitomo Electric for fiber-optic equipment [citation:3][citation:12].

Key details:

Meta's internal memo contains a candid admission: adopting the latest GPUs at Meta's scale "has been a heavy lift, and it has cost us time" [citation:12]. Iris is designed to run ranking and recommendation systems (which power Facebook and Instagram feeds) and generative AI features — alongside, not instead of, the GPUs Meta continues to buy from Nvidia and AMD [citation:12].

The chip effort sits inside a much larger infrastructure push. Meta plans to double its data center computing capacity from 7 gigawatts in 2026 to 14 gigawatts in 2027. One gigawatt is roughly enough to power 800,000 homes; 14 gigawatts would power more than 11 million [citation:12].

Iris is Broadcom's third major custom chip project after Google's TPU and OpenAI's Jalapeño — a sign that the chip design firm has become the go-to partner for companies that want custom silicon without building a chip design team from scratch [citation:12].

Why Now? The Economics of Inference

The timing of these announcements is no coincidence. All four companies are responding to the same structural shift: the AI industry's compute center of gravity is moving from training to inference.

Training is a one-time cost. You train a model once (or a few times per year). It's expensive, but it's a capital project.

Inference is a recurring cost. Every user query, every API call, every agent action consumes compute. As products scale, inference costs become an operational expense that can quickly eclipse training costs.

DeepSeek's situation illustrates the math. With tens of millions of daily active users, each query costing a fraction of a cent, the aggregate bill becomes massive. DeepSeek's V4 Flash is priced at 2 yuan per million output tokens — aggressively low to capture market share. But that strategy only works if you can bring down the underlying compute cost to match [citation:1].

Anthropic is running a similar calculation. With annualized revenue run rate exceeding $47 billion and a growth curve that's nearly vertical, the company needs to squeeze every dollar of efficiency out of its compute stack [citation:15].

For Meta, the pressure comes from volume. Its recommendation systems power billions of users, and adding generative AI features multiplies the compute demand further [citation:12].

For OpenAI, the math is about both cost and control. As ChatGPT, Codex, and agentic products scale, having a chip designed specifically for the workloads it actually runs becomes a competitive advantage [citation:4].

The Bigger Picture: From Buying to Defining Compute

The deeper story is not about individual chips but about a structural power shift in the AI industry.

For the past decade, the dominant model has been simple: chip companies (primarily Nvidia) design general-purpose processors; AI companies buy them. The chip companies defined the compute; the AI companies built on top of it.

That model is now being inverted. AI companies are beginning to define the compute they need and telling the industry how to build it.

This manifests in several ways:

Broadcom has emerged as a key enabler of this shift. The chip design firm is the common thread across three major custom silicon programs: Google's TPU, OpenAI's Jalapeño, and Meta's Iris [citation:12]. Analysts estimate Broadcom's AI-related revenue reached $10.8 billion in its fiscal second quarter of 2026, up 143% year-over-year, and the company has forecast AI revenue exceeding $100 billion in fiscal 2027 [citation:13].

What is notable is that all these chips — regardless of who designs them — are manufactured by TSMC. This means the AI chip design battle is fought on software and architecture, not manufacturing. Nvidia's dominance is not threatened by manufacturing constraints but by the fact that its customers are increasingly capable of designing chips that do exactly what they need, and nothing more.

For now, Nvidia remains the dominant player, with approximately 74% of the inference market share [citation:15]. But the trend lines are clear: AI companies are building their own chips not to replace Nvidia today but to define their own compute destiny for tomorrow.

Key Takeaways

Shop Smart Wearables at Gzmato

Whether you need audio glasses, camera-first glasses, a privacy-conscious HUD, or a full AR experience for gaming and media, Gzmato carries the full range of smart wearables.