The “honeymoon phase” between AI companies and NVIDIA appears to be ending.

On July 2, The Information reported that Anthropic is in talks with Samsung to develop custom AI chips. It's been barely a week since OpenAI unveiled its first custom inference chip, Jalapeño (fabricated by Broadcom). Add Google's TPU and Amazon's Trainium to the mix, and four of the world's top five AI companies are now advancing self-developed chips. NVIDIA's monopoly in AI training chips is being eroded from the supply side.

Why Is Everyone Building Chips?

Anthropic's demand for chips is real and urgent. In June, demand exploded after Claude Sonnet 5 launched, and compute requirements for the Mythos and Fable models are astronomical. Although Anthropic has secured compute commitments from Amazon and Google, in a market where NVIDIA H100/B100 chips remain supply-constrained, putting your core compute lifeline entirely in third-party hands is too risky.

The logic for custom chips is simple: general-purpose GPUs (like NVIDIA's product line) sacrifice efficiency for flexibility, while chips designed specifically for inference or particular model architectures can deliver order-of-magnitude advantages in performance per watt. OpenAI claims Jalapeño surpasses competitors on inference efficiency; Google's TPU is already on its fifth generation, powering training and inference for the entire Gemini model lineup; Amazon's Trainium and Inferentia are deployed at scale within AWS.

Choosing Samsung as a partner is also deliberate. Samsung itself is a key NVIDIA fabrication partner with deep expertise in HBM (High Bandwidth Memory) and advanced packaging. Samsung also collaborates with Google on chip fabrication and has extensive foundry experience. For Anthropic, partnering with Samsung leverages Samsung's manufacturing capabilities without facing the geopolitical risks that come with relying on TSMC.

The Compute Sovereignty Arms Race

WAIC 2026 is about to open, and QbitAI's preview identified this year's core question in compute: can super-nodes and optical interconnects bypass the single-chip physical ceiling? As single-chip process progress slows and the power wall grows harder to break, connecting multiple chips via high-speed interconnect into “super-nodes” is becoming a new direction. But this only delays the problem — it doesn't solve it.

The deeper driver is cost. When an AI company spends billions or even tens of billions annually on compute, even a 10% optimization in chip costs means hundreds of millions in savings. AI chip startups like Etched AI have already reached valuations around $5 billion, indicating capital markets see opportunity in AI chip customization.

Of course, building your own chips is far from easy. Anthropic itself acknowledges that a diversified hardware strategy — including Google, Amazon, and Nvidia chips — will remain central to its compute strategy. Custom chips won't replace NVIDIA overnight; they're more of an insurance policy and long-term bet. But the trend is clear: top AI companies don't want to be bottlenecked by any single chip supplier.

NVIDIA won't sit idle. Its moats in the CUDA ecosystem, software stack, and developer community remain deep. But when customers start building their own chips, NVIDIA must prove: continuing to use my chips is more cost-effective than building your own. The next stage of AI chip competition won't be single-chip performance comparisons — it'll be a battle over the efficiency, cost, and ecosystem of entire hardware-software systems.