Who makes
the silicon.
The previous page covered what the chips are — GPU, FPGA, ASIC. This one covers who actually builds them, what each company is shipping right now, and where the race really stands. Not the hype version.
Verified 12 Aug 2026
This is a snapshot of a market that moves every quarter. Figures are the most recent published by the companies themselves; sources at the bottom. Share prices refresh each trading day — last update 2026-08-12.
One giant, and everyone else
Almost every serious AI model in the world was trained on NVIDIA hardware, and most still run inference on it. That has been true for a decade and it is still true today. But the interesting part of 2026 isn't NVIDIA's lead — it's who finally has a real reason to build an alternative.
Two things changed. First, the customers got big enough to build their own chips: when you're spending tens of billions a year on compute, designing your own silicon stops being exotic and starts being obvious accounting. Second, the workload split. Training a frontier model is a brutal, flexible, once-in-a-while job that rewards NVIDIA's generality. Inference — actually answering user questions, forever — is repetitive and predictable, which is exactly the shape of problem a fixed-function ASIC was invented for.
NVIDIA still wins training. The fight worth watching is inference, because that's the bill that never stops arriving — and it's the workload a custom chip can actually beat a GPU at.
Everyone who matters
NVIDIA
GPU · the incumbent
NVDA $217.50 ▼0.0%
Sells to everyone · Blackwell → Vera Rubin
The default, by an enormous margin. NVIDIA's advantage was never only the chips — it's CUDA, the software layer every AI framework was built against over fifteen years. A competitor can match the silicon and still lose, because the customer's code doesn't run. That moat is the single most underrated fact in the industry.
The current generation is Blackwell. The next is Vera Rubin — the Rubin R100 GPU paired with a custom 88-core Vera CPU — announced at the start of 2026 with full production shipments targeted for the second half of the year. NVIDIA claims roughly 5× Blackwell's inference performance; treat vendor performance claims as marketing until independent numbers land.
- $75.2BData-center revenue in a single quarter (Q1 FY2027, ended 26 Apr 2026)
- +92%Year-over-year growth in that number
- $193.7BFull-year data-center revenue, FY2026
AMD
GPU · the challenger
AMD $474.32 ▲1.0%
Instinct MI450 · Helios rack · ROCm
The only company selling a general-purpose AI GPU that competes head-on. AMD's problem was never the hardware — it was ROCm, the software stack that has to stand in for CUDA. That gap has narrowed enough that the biggest buyers in the world are now willing to bet on it.
And they have. OpenAI committed to 6 gigawatts of AMD GPUs, with the first gigawatt of MI450 deploying in the second half of 2026. The deal even includes stock: OpenAI earns the right to buy up to 160 million AMD shares for a cent apiece as its deployments hit milestones — a measure of how badly AMD wanted the win. Meta signed a matching 6 GW agreement, and Anthropic committed to up to 2 GW. The MI450 ships with 432 GB of HBM4 — the ultra-fast stacked memory every AI chip is starved for — in AMD's Helios rack, positioned directly against NVIDIA's rack-scale systems.
Because the deals got too big to count in chips. A gigawatt is a power plant's worth of electricity — roughly one nuclear reactor running flat out. "6 GW of GPUs" means "however many hundreds of thousands of chips it takes to draw that much power." When you see gigawatts in a headline, read it as city-sized amounts of computing.
- 6 GWCommitted by OpenAI, and separately by Meta
- 2 GWCommitted by Anthropic
- 432 GBHBM4 memory per MI450
ASIC · TPU
GOOGL $343.80 ▼3.8%
TPU v7 "Ironwood" · designed with Broadcom
Google has been building its own AI chip since 2015, long before it was fashionable, and the TPU is the most mature custom AI silicon on earth. Gemini is trained and served on it. The current generation, Ironwood (TPU v7), is generally available in Google Cloud, and the eighth generation is set to split into separate training and inference parts — an admission that those two jobs now deserve different chips.
The strategic shift is that Google is no longer keeping TPUs to itself. Anthropic's April 2026 agreement with Google and Broadcom covers multiple gigawatts of next-generation TPU capacity coming online from 2027 — the clearest signal yet that TPUs are a product, not just an internal cost saving.
- 2015First TPU deployed internally — a decade of head start
- v7Current generation, "Ironwood", GA in Google Cloud
- Multi-GWAnthropic capacity commitment, from 2027
Amazon / AWS
ASIC · Trainium
AMZN $272.27 ▼2.1%
Trainium2 & Trainium3 · Annapurna Labs
AWS designs Trainium (training) and Inferentia (inference) in-house through Annapurna Labs. The flagship proof point is Project Rainier — a cluster of roughly half a million Trainium2 chips in Indiana, built for Anthropic, which came online in late 2025 and has since grown past a million chips. Trainium3 followed at re:Invent 2025, manufactured on a cutting-edge 3 nm process at TSMC — the Taiwanese foundry that fabricates nearly every advanced chip on this page.
The logic is the same as Google's: AWS doesn't need Trainium to beat NVIDIA in benchmarks. It needs it to be cheaper per token inside AWS — and to give Amazon leverage in every future NVIDIA negotiation.
- 1M+Trainium2 chips running Anthropic's Claude
- ~500KChips in Project Rainier at activation
- 3 nmTSMC process for Trainium3
Broadcom
ASIC · the arms dealer
AVGO $416.08 ▼1.5%
Custom silicon for other people's names
The most important AI chip company most people have never heard of. Broadcom doesn't sell a branded accelerator — it co-designs and builds the custom chips that other companies put their own name on, Google's TPU among them. When a hyperscaler announces "our own AI chip," Broadcom or Marvell is very often the engineering behind it.
That quiet position has become one of the largest AI businesses in existence. Broadcom has guided to roughly $56 billion of AI semiconductor revenue in fiscal 2026, and says it has line of sight to more than $100 billion in 2027 against a large committed backlog.
- ~$56BGuided FY2026 AI semiconductor revenue
- >$100BStated line of sight for 2027
- 0Consumer-recognised chip brands
SK Hynix · Samsung · Micron
HBM · the memory bottleneck
High-bandwidth memory · the part every chip above fights over
None of the processors on this page work without HBM — high-bandwidth memory, stacked in towers directly beside the chip so data arrives fast enough to keep thousands of cores fed. Exactly three companies on earth make it: SK Hynix, Samsung and Micron. That makes memory, not processors, the quiet chokepoint of the whole industry — how many GPUs get built is routinely limited by how much HBM exists to put on them.
SK Hynix is the leader and NVIDIA's principal supplier, a position that turned a boom-and-bust commodity business into some of the most sought-after capacity in tech. The newest generation, HBM4, is what the MI450 and Vera Rubin era is built around — it's the "432 GB" in AMD's spec sheet above.
- ₩1,504,000 ▲5.5%SK Hynix (KRX: 000660) — the HBM leader
- $868.52 ▲0.9%Micron (NASDAQ: MU) — the American maker
- ₩255,500 ▲6.7%Samsung (KRX: 005930) — the giant playing catch-up
Cerebras
Wafer-scale · inference speed
CBRS $230.01 ▲1.4%
CS-3 · the entire wafer is the chip
The most physically radical design in the industry. Normal manufacturing cuts a silicon wafer into hundreds of small chips; Cerebras doesn't cut it — one wafer becomes one enormous processor. That removes the slow hop between chips, which is what makes it extraordinarily fast at generating tokens. The market noticed: Cerebras went public in May 2026 in the year's biggest tech IPO, pricing at $185 a share and nearly doubling on its first day.
In March 2026 AWS partnered with Cerebras to serve inference on Amazon Bedrock, splitting the job in two: Trainium reads your prompt, Cerebras writes the answer — each chip doing the half it's best at. A specialist being adopted by a hyperscaler that builds its own silicon is a meaningful endorsement.
Groq
LPU · inference only
PRIVATE
Deterministic low-latency token generation
Groq builds an LPU (Language Processing Unit), an inference-only chip that runs like a train timetable: every operation is scheduled in advance, nothing waits and nothing guesses, so answers start fast and arrive at a steady, predictable pace. It doesn't train models and doesn't try to. It's a clean example of the thesis running through this whole page: as inference becomes the dominant cost, hardware built for only inference starts to make commercial sense. Still private, it raised another $650 million in mid-2026.
Intel
The struggle
INTC $97.71 ▲0.2%
Gaudi accelerators · foundry ambitions
The company that defined computing for thirty years has not landed a punch in AI. Its Gaudi accelerators never achieved meaningful share against NVIDIA, and its more consequential bet is arguably the foundry — manufacturing chips for other people, in a market where nearly everything leading-edge is made by TSMC. Intel remains dominant in ordinary server CPUs; that is simply not where the AI money is.
Huawei
China · Ascend
NOT LISTED
Ascend 910C · 950 series planned
US export controls bar NVIDIA's best chips from China, which forced a domestic alternative into existence. Huawei's Ascend line is the result and it is genuinely deployed at scale inside China. It is also, by Huawei's own published roadmap, still behind: the Ascend 950 parts planned for 2026 carry lower headline throughput than the existing 910C, and the whole programme is constrained by manufacturing access rather than design talent.
Worth understanding as geopolitics rather than a product race — the constraint is which fabrication equipment China can legally buy, and that changes with policy, not engineering.
Apple & Qualcomm
NPU · on-device
AAPL $304.91 ▼1.1%
QCOM $162.68 ▲0.3%
The chip already in your pocket
A completely different market that rarely shares a headline with the others. Every recent iPhone, Mac, and Android flagship contains an NPU — a small ASIC-style block for running models locally, with no data centre involved. Nobody is training a frontier model on one, but they are almost certainly the most numerous AI chips on the planet.
Apple is a quiet outlier in a second way: its unified memory architecture lets a Mac Studio address far more memory than a discrete GPU of similar price, which is exactly why a Mac can run very large models that a gaming card physically cannot. We cover that on the local AI hardware page — it's the single most counterintuitive fact in the whole subject.
Three strategies, not ten companies
Strip away the brand names and there are only three plays being run:
NVIDIA, AMD
Build the best general-purpose GPU and sell it to everyone. Highest volume, highest margin, hardest to displace — the software ecosystem is the real product.
Google, Amazon, Meta, OpenAI
Design custom ASICs for your own workloads, usually with Broadcom or Marvell. Doesn't need to win benchmarks — only needs to be cheaper for you, and to give you leverage.
Cerebras, Groq
Do one thing — fast inference — better than a general chip can. Small share, real technical advantage, and increasingly picked up by the giants.
Notice what the second column means. Google, Amazon and Meta are simultaneously NVIDIA's largest customers and its most serious long-term competitors. That's an unusual and unstable arrangement, and it's the thing most likely to reshape this market.
You can't buy most of these
Here's the part the coverage usually skips. Of everything above, a normal business can only actually buy NVIDIA and AMD. TPUs and Trainium aren't for sale at any price — you rent them, inside Google Cloud and AWS respectively. Custom ASICs are built by companies at hyperscale, for themselves. Cerebras and Groq you reach through an API.
So the practical decision is much smaller than the news makes it look: an NVIDIA GPU rented by the hour, or an NVIDIA GPU in a rack in your office. And the question that decides it isn't brand — it's how much memory you need, which is covered on the local AI hardware page, and whether your data can leave the building, which is frontier vs local.
We run models on our own hardware in Las Vegas — a mix of NVIDIA accelerators and Apple Silicon — precisely because the economics above are real. When inference is the bill that never stops, owning the machine changes the maths. That's the same reasoning driving every company on this page.
Common questions
Who makes AI chips?
NVIDIA makes the GPUs that run most AI and remains far ahead. AMD is the main GPU alternative. Google (TPU), Amazon (Trainium/Inferentia), Meta, Microsoft and OpenAI design custom ASICs, usually engineered with Broadcom or Marvell. Cerebras and Groq build specialist inference hardware. Intel and Huawei compete at the margins. Apple and Qualcomm build the NPUs in phones and laptops.
Does anyone actually compete with NVIDIA?
Yes, but unevenly. NVIDIA still takes the large majority of data-center AI revenue and CUDA is the real moat. Pressure comes from two directions: AMD, which won multi-gigawatt commitments from OpenAI, Meta and Anthropic for MI450; and the hyperscalers' own chips, which don't have to beat NVIDIA on paper — only to be cheaper for their owner's own workloads.
What is Broadcom's role in AI chips?
Broadcom doesn't sell a branded AI chip — it co-designs and builds the custom accelerators other companies put their name on, including Google's TPU. That makes it one of the largest AI chip businesses in the world while staying nearly invisible publicly. It has guided to roughly $56 billion of AI semiconductor revenue in fiscal 2026.
Which AI chip should my business buy?
Almost certainly none of the famous ones. TPUs and Trainium can't be bought at all — they're rented inside Google Cloud and AWS. Custom ASICs are built by hyperscalers for themselves. For a normal business the real choice is an NVIDIA GPU, rented or owned, and the decision that actually matters is how much VRAM you need.
Sources. Company figures are taken from primary disclosures where available: NVIDIA Q1 FY2027 results and Q4 FY2026 8-K; OpenAI–AMD partnership and AMD–Meta announcement; Anthropic–Google–Broadcom compute partnership; AWS Project Rainier and AWS–Cerebras collaboration. Market-share percentages are deliberately omitted: published estimates vary widely and none are auditable. Revenue figures reported by the companies themselves are the honest way to see the gap.