Published: July 11, 2026 | Reading Time: ~14 minutes | Channel: techminute
Here's a number that should keep Jensen Huang up at night: in the span of just sixteen days between June 24 and July 9, 2026, four of the world's most important AI companies — OpenAI, Anthropic, DeepSeek, and Meta — each took concrete steps toward building their own AI silicon. Not rumors. Not "exploring options." Actual chips entering production, partnerships with Broadcom and TSMC, engineers being hired, internal memos leaked.
This isn't a trend. It's an exodus.
NVIDIA still owns the AI training market — nobody's seriously challenging the H200 and its successors on raw training throughput. But training isn't where the money is going anymore. Inference is. And inference, it turns out, is a fundamentally different problem that doesn't need a $40,000 GPU to solve. Every major AI lab has reached the same conclusion: why pay NVIDIA's 70% margins when you can build exactly what your models need for a fraction of the cost?
What follows is the story of how the AI industry's biggest customers became NVIDIA's newest competitors — all at the same time.
To understand why this is happening now, you need to understand the shifting economics of AI compute.
When ChatGPT launched in late 2022, the bottleneck was training — building the model in the first place. Training a frontier model requires massive clusters of GPUs running in parallel for weeks or months, communicating constantly through high-bandwidth interconnects. NVIDIA's CUDA ecosystem and NVLink fabric are genuinely unmatched here, and nobody pretends otherwise.
But something changed in 2025 and accelerated dramatically in 2026: inference surpassed training as the dominant cost center. Every time someone asks ChatGPT a question, runs a Codex agent on a bug, or generates an image with Gemini Omni, that's inference. And as AI products have gone from "novelty demos" to "things billions of people use daily," the inference bill has exploded.
Here's the key insight: inference workloads are fundamentally different from training workloads. They're latency-sensitive, not bandwidth-sensitive. They benefit more from specialized architectures optimized for specific model families than from general-purpose GPU muscle. And the market for inference chips is vastly larger — because while you train a model once, you run inference on it millions or billions of times.
OpenAI's Greg Brockman put it plainly on the company's podcast last year: "We have a deep understanding of the workload. We've really been looking for specific workloads that are underserved, and asking how can we build something that will accelerate what's possible?"
Translation: we know exactly what our models need, and NVIDIA's one-size-fits-all approach is leaving performance on the table.
What makes this story remarkable isn't just that these companies are building chips — it's how differently they're approaching the problem, reflecting their unique positions in the AI ecosystem.
OpenAI was first to blink. On June 24, the company unveiled "Jalapeño," a custom inference processor built with Broadcom. It's an inference-only play — OpenAI is explicitly not trying to replace NVIDIA for training workloads.
The chip was designed with direct input from OpenAI's own models, a recursive twist that feels appropriate for 2026. Early testing shows "significantly better performance-per-watt than current state-of-the-art alternatives," according to the company. The partnership with Broadcom was first announced in October 2025, but this was the formal unveiling.
What's most revealing is how OpenAI framed it: "OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience."
That's not a chip announcement. That's a vertical integration manifesto.
Critically, Jalapeño was designed specifically around the inference patterns of real-time coding models — Codex, in other words. That's OpenAI's highest-volume, most latency-sensitive workload. They're not trying to build a chip for everything. They're building a chip for the one thing that costs them the most money.
Just eight days later, The Information reported that Anthropic was in talks with Samsung about a custom chip collaboration. Compared to OpenAI's brash launch, Anthropic's approach is characteristically measured.
According to the TechCrunch report, Anthropic "hasn't yet decided what the chip will be used for, how it will fit into the server, or how powerful it will be." The company emphasized that "a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy."
But don't mistake caution for lack of ambition. The Samsung angle is particularly interesting because of the manufacturing capability it implies. Samsung has been talking about 2nm process technology, and is already a major NVIDIA manufacturing partner. If Anthropic can leverage Samsung's advanced nodes — and Samsung's eagerness to diversify beyond NVIDIA — this could become the most technologically advanced of the four chips.
It's also worth noting the timing: Anthropic's news came roughly one week after OpenAI's announcement. That's not a coincidence. When your chief competitor unveils a chip that could give them a 50% inference cost advantage, "we're thinking about it" stops being an acceptable answer.
On July 7, Reuters dropped a bombshell: DeepSeek, the Chinese AI lab that shocked the world with its cost-efficient models, is building its own AI inference chip.
Three sources told Reuters that DeepSeek is already in talks with manufacturing partners and quietly hiring chip engineers. The goal: reduce reliance on both NVIDIA and Huawei hardware for inference.
This one hits differently because of DeepSeek's track record. This is the company that built competitive frontier models at a fraction of what US labs were spending. If anyone can figure out how to build inference chips that punch above their weight class, it's DeepSeek. Analyst Richard Windsor of Radio Free Mobile was blunt: "Nvidia is at zero in China and staying there. DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading edge manufacturing."
The export control angle adds a layer of geopolitical complexity the other three don't face. DeepSeek's chip, if successful, would primarily serve the domestic Chinese market — but that market is enormous, and if DeepSeek can run inference at dramatically lower costs than competitors using sanctioned NVIDIA chips, it reshapes the competitive landscape in Asia overnight.
Then came Meta, and Meta doesn't do anything small.
An internal memo obtained by Reuters revealed that Meta's latest MTIA (Meta Training and Inference Accelerator) chip — reportedly code-named "Iris" — will enter production at TSMC in September 2026. One chip variant "sailed through its testing phase in about six weeks," the memo said.
The scope here dwarfs the others. Meta plans to deploy 7 gigawatts of compute this year, and double that next year. Capital expenditure for 2026 is projected at $125–145 billion. The supply chain is already locked in: Broadcom for design, TSMC for manufacturing, Samsung for RAM, Sandisk for storage, Sumitomo Electric for fiber-optic interconnects.
But here's what makes Meta's play fundamentally different: they're not just building inference chips. The MTIA program, running since 2023, targets both training and inference — specifically for Meta's ranking and recommendation algorithms, then broader AI workloads, and eventually inference for consumer-facing applications.
Meta is also hedging aggressively. Even as they build their own silicon, they've signed multi-billion-dollar deals with AMD for Instinct GPUs, with Amazon for homegrown CPUs, and with ARM for recommendation system compute. This isn't "replace NVIDIA." This is "build optionality into every layer of the stack."
| Player | Chip Name | Partner | Purpose | Status | Timeline |
|---|---|---|---|---|---|
| OpenAI | Jalapeño | Broadcom | Inference (coding models) | Testing | Announced June 24 |
| Anthropic | TBD | Samsung (talks) | TBD | Early discussions | Reported July 2 |
| DeepSeek | TBD | TBD | Inference | Hiring/manufacturer talks | Reported July 7 |
| Meta | Iris (MTIA) | Broadcom + TSMC | Training + Inference | Entering production | September 2026 |
| Metric | NVIDIA H200 | OpenAI Jalapeño (claimed) |
|---|---|---|
| Performance-per-watt | Baseline | "Significantly better" |
| Cost savings vs GPU inference | — | ~50% (early testing) |
| Target workload | General-purpose | Model-specific (Codex) |
The second-order effects of this shift are going to ripple through the industry for years.
NVIDIA's moat is shrinking — but not where you think. The training market remains secure. The H200 and its successors are still the only serious option for pre-training frontier models, and CUDA's software ecosystem is decades deep. But inference is the growth market, and NVIDIA's share there is now under direct assault from its own customers.
TSMC becomes the true bottleneck. When all four companies — plus Google, Amazon, and Microsoft — are competing for advanced packaging and leading-edge node capacity, TSMC's allocation decisions become geopolitically significant. Whoever gets the wafers wins.
The "chipflation" era may be peaking. Meta's memo explicitly references "an unprecedented component shortage." But if these custom chips succeed, the demand for NVIDIA's premium-priced GPUs for inference workloads drops. That could actually ease supply pressure — or it could just shift the bottleneck to TSMC.
Custom silicon becomes table stakes. If you're a serious AI company in 2026 and you don't have a chip program, you're now behind. Expect Apple, Microsoft, and Amazon to accelerate their existing silicon efforts. Expect startups like Groq, Cerebras, and d-Matrix to face both opportunity (everyone wants alternatives) and threat (their potential customers are becoming competitors).
The open-source chip ecosystem gets interesting. If DeepSeek open-sources aspects of its chip design — as it did with its models — it could spark a RISC-V-style revolution in AI accelerators. Unlikely in the near term given export controls, but worth watching.
Let's be honest about what this story is — and isn't.
These chips won't ship tomorrow. Meta's September production start means chips in data centers in 2027 at the earliest. OpenAI's Jalapeño is still in testing. Anthropic hasn't even decided what to build. DeepSeek is still hiring. This is a structural shift, not a near-term NVIDIA killer.
Custom chips are hard — really hard. Google's TPU program took years to become competitive. Amazon's Trainium had a rocky start. Building a chip that actually outperforms NVIDIA's latest on real workloads, not just spec sheets, is an enormous engineering challenge. Some of these projects will fail or underdeliver.
NVIDIA isn't standing still. The company has its own inference-optimized products (L40S, upcoming Rubin architecture) and deep relationships with every hyperscaler. They've seen this coming and are responding.
Software is the real moat. CUDA's ecosystem — libraries, frameworks, developer familiarity — took 15+ years to build. Custom chips need custom software stacks, and that's often harder than the silicon itself. OpenAI has the in-house talent. Meta has the engineering muscle. Anthropic and DeepSeek? Less clear.
Geopolitics could scramble everything. If TSMC capacity becomes a political football between the US and China, DeepSeek gets locked out. If Samsung's 2nm yields disappoint, Anthropic's timeline slips. The semiconductor supply chain is fragile in ways software people don't always appreciate.
In sixteen days, the AI industry crossed a Rubicon. The companies that build the models have decided they also need to build the chips that run them. Inference — the boring, unglamorous, "just running the model" part of AI — has become the strategic battleground, and everyone wants their own weapons.
NVIDIA isn't going anywhere. But the era where "buy more GPUs" was the answer to every AI scaling question is over. The next chapter is being written in silicon, one custom chip at a time.
[TechCrunch] — OpenAI unveils its first custom chip "Jalapeño," built by Broadcom. Russell Brandom, June 24, 2026. https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom/
[TechCrunch] — Anthropic is discussing a new custom chip with Samsung. Lucas Ropek, July 2, 2026. https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/
[Engadget / Reuters] — DeepSeek is developing its own AI chips. Daniel Cooper, July 7, 2026. https://www.engadget.com/2209378/deepseek-reportedly-developing-ai-chips/
[TechCrunch] — Meta's new AI chips will begin production in September. Ram Iyer, July 9, 2026. https://techcrunch.com/2026/07/09/metas-new-ai-chips-will-begin-production-in-september/
[Android Headlines] — Meta Is Massively Ramping Up Its In-House AI Silicon Production This September to Fight 'Chipflation.' July 2026. https://www.androidheadlines.com/2026/07/meta-iris-ai-chip-production-september-custom-silicon.html
[Reuters / US News] — Exclusive: China's DeepSeek Developing Its Own AI Chip, Sources Say. July 7, 2026. https://www.usnews.com/news/top-news/articles/2026-07-07/exclusive-chinas-deepseek-developing-its-own-ai-chip-sources-say
All claims verified against Gold-tier (official announcements, Reuters original reporting, internal memos) and Silver-tier (TechCrunch, Engadget) sources. Each source URL was scraped and confirmed accessible. Last verified: July 11, 2026.
