CompetitorsWide moat
Nvidia (NVDA) — moat facet
The question isn't who competes with Nvidia but from which direction — and where the moat was thinnest, in inference, Nvidia paid $20 billion to license the threat rather than argue with it.
Nvidia holds something close to 70% of the AI-chip market1, which makes the interesting question not who competes with it but from which direction. Four attacks are live, and only one of them is a company selling the same product to the same buyers.
That one is AMD, the only other merchant GPU maker, now shipping rack-scale systems and holding a six-gigawatt commitment from OpenAI2. The more structural attack comes from Broadcom, which does not sell chips against Nvidia so much as help Nvidia's own customers design their own: six named XPU customers, seven generations of Google TPU since 2014, and a ten-gigawatt OpenAI programme3. Custom ASIC shipments are forecast to grow far faster than merchant GPUs4, which is the number that should worry Nvidia most, and the reason those buyers appear again as a root threat on this company's own list.
The third front is geographic and was not chosen by anyone at Nvidia. Export controls removed it from China, and Huawei filled the space: Nvidia's share of Chinese AI silicon has collapsed from near-total toward a single-digit forecast while Huawei's Ascend line scales into the gap5. A market that once looked like Nvidia's next decade now belongs to a rival that American policy created.
The fourth is the most revealing, because of how Nvidia answered it. Inference — running trained models rather than building them — is where its GPUs are least efficient and where specialist chips had the best case. In December Nvidia paid about $20 billion for the assets and technology of Groq, the leading inference challenger6, through a licence rather than a takeover. That is the competitive posture in one transaction: where the moat is thinnest, Nvidia would rather own the threat than argue with it. The pages that follow take the four in that order.
Nvidia's roughly 70% share has held through a year in which every rival got better funded: AMD won a six-gigawatt commitment, Broadcom's XPU roster grew, and inference startups raised records. Holding rather than widening — the position is intact, but for the first time each attack vector has a credible, financed champion behind it.
Four attacks are live from different directions — a merchant rival, a custom-silicon arms dealer, a state-created champion in China, and inference specialists. Watch the share itself: it has survived a year in which every rival got better funded, and its erosion would show up here before it shows up in revenue.
Source: Custom-ASIC market coverage (Tom's Hardware) ↗- Third-party estimateNvidia holds roughly 70% of the AI-chip market.Tom's Hardware — custom AI ASIC state of play: Broadcom with six named XPU customers, seven generations of co-designed Google TPU since 2014, OpenAI's 10 GW programme with first deployment targeted 2H 2026; custom ASIC shipments forecast to grow far faster than merchant GPUs — 2026 · publ. May 2026 · source ↗
- ReportedAMD holds a six-gigawatt multi-generation Instinct commitment from OpenAI.AMD press release — AMD and OpenAI strategic partnership: 6 gigawatts of Instinct GPUs across generations, first 1 GW of MI450 in 2H 2026; AMD expects tens of billions of dollars in revenue — October 2025 · publ. October 6, 2025 · source ↗
- ReportedBroadcom has six named XPU customers, seven generations of Google TPU since 2014, and a 10 GW OpenAI programme.Tom's Hardware — custom AI ASIC state of play: Broadcom with six named XPU customers, seven generations of co-designed Google TPU since 2014, OpenAI's 10 GW programme with first deployment targeted 2H 2026; custom ASIC shipments forecast to grow far faster than merchant GPUs — 2026 · publ. May 2026 · source ↗
- Third-party estimateCustom ASIC shipments are forecast to grow far faster than merchant GPUs.Tom's Hardware — custom AI ASIC state of play: Broadcom with six named XPU customers, seven generations of co-designed Google TPU since 2014, OpenAI's 10 GW programme with first deployment targeted 2H 2026; custom ASIC shipments forecast to grow far faster than merchant GPUs — 2026 · publ. May 2026 · source ↗
- Third-party estimateNvidia's Chinese AI-silicon share has collapsed toward single digits as Huawei's Ascend scaled.Market coverage of China AI silicon — NVIDIA's share of the Chinese AI-chip market collapsing from near-total toward a single-digit forecast as export controls took hold, with Huawei's Ascend line scaling into roughly half or more of the market; Chinese frontier models reported as optimised for Ascend hardware — 2026 · publ. 2026 · source ↗
- ReportedNVIDIA paid about $20 billion for the assets and technology of Groq, the leading inference challenger.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 ↗
- Nvidia Form 10-K filings — Business & Risk Factors (SEC EDGAR)
- Custom AI ASIC state of play (Tom's Hardware)