The Inference Challengers, One of Which Nvidia BoughtNarrow moat
Nvidia (NVDA) — moat facet
Nvidia's defence was never that no better chip exists — only that no rival ecosystem does. Groq cost $20 billion to keep it that way.
Inference is the part of AI that keeps growing and the part Nvidia is least suited to. Training rewards raw parallel throughput, which is what a GPU is; serving a trained model to millions of users rewards latency and cost per token, where purpose-built chips have a real architectural case. With inference expected to take the larger share of AI compute spending, that is the widest crack in the franchise.
Nvidia's response was not a product. In December it paid roughly $20 billion for the assets of Groq, the best-known inference-chip challenger1, structured as a non-exclusive licence to Groq's technology; Nvidia's annual report now lists the arrangement as a risk, warning that it may be unable to recover the associated costs4. It is a variant of the Mellanox manoeuvre: identify the adjacent technology that could commoditise you, and secure it while it is still cheaper than the damage.
The field did not empty out. Cerebras went public in May 2026 and OpenAI committed some $20 billion to its chips2, while AI-chip startups collectively raised billions more during the year3. The pattern to watch is whether any of them wins a second large customer — a single anchor buyer proves the architecture but not the business, and Nvidia's real defence has never been that no better chip exists, only that no rival ecosystem does. An inference startup that lands three hyperscalers would be the first genuine counter-example.
Paying about $20 billion for Groq's assets and a licence to its technology put the most credible architectural argument against GPUs inside Nvidia's roadmap — a variant of the Mellanox manoeuvre, and a genuine strengthening of the weakest part of the franchise. What keeps it from being decisive is that the licence is non-exclusive and the field refilled immediately: Cerebras went public with a $20 billion OpenAI commitment behind it.
Buying Groq removed one challenger; the money keeps creating new ones. Funding still rising after the Groq deal says investors believe inference can be won from Nvidia.
Source: CNBC, citing Dealroom (Apr 2026) ↗- ReportedNVIDIA acquired assets from inference-chip startup Groq for about $20 billion.CNBC, 'Nvidia AI chip rivals attract record funding as competition heats up' - in December NVIDIA acquired assets from AI inference startup Groq for $20 billion; AI-chip startups raised $8.3 billion globally in 2026 by mid-April, per Dealroom — December 2025 - April 2026 · publ. 2026-04-17 · source ↗
- ReportedCerebras went public in May 2026 with OpenAI committing some $20 billion to its chips.DigiTimes — OpenAI's ~$20 billion commitment to Cerebras chips alongside NVIDIA's ~$20 billion Groq deal, as inference becomes the larger share of AI compute spending — April 2026 · publ. April 2026 · source ↗
- ReportedAI-chip startups attracted record funding through 2026.CNBC — AI-chip startups challenging NVIDIA attracting record funding through 2026 — 2026 · publ. April 2026 · source ↗
- ReportedNVIDIA's 10-K describes the Groq deal as an intellectual-property licence and warns it may not recover the costs.NVIDIA Form 10-K (FY2026), risk factors - 'We have entered into an intellectual property license arrangement with Groq, Inc., or Groq, that required significant, nonrefundable payments' and 'we may be unable to recover the associated costs or realize an adequate return on this spend'; the Q2 FY2027 10-Q describes it as the Groq, Inc. non-exclusive license agreement — FY2026 · publ. Filed Feb 2026 · source ↗
- Nvidia Form 10-K filings — Business & Risk Factors (SEC EDGAR)
- Nvidia-Groq acquisition (Forbes)
- AI-chip rivals attract record funding (CNBC)