Originally published on The Daily Flare.
AI chip demand is breaking records in 2026: Nvidia just posted $75.2 billion in data-center revenue in a single quarter. Here is why these processors are becoming the most strategically important commodity in tech.
AI chips are processors built to do the massive parallel math behind training and running AI models, and they handle it orders of magnitude better than general-purpose CPUs. What changed is not how they work — it is how badly the world suddenly needs them, and how much money is chasing them.
How AI chip demand rewrote the market forecasts
Start with the market math. Precedence Research estimates the AI chip market at $94.44 billion in 2025, growing to $121.73 billion in 2026 and reaching $1.10 trillion by 2035, at roughly 27.9% annual growth. Another firm, SNS Insider, puts the 2035 figure closer to $1.35 trillion. The estimates disagree by roughly 2x because the firms define “AI chip” differently — data-center GPUs and custom accelerators only, versus every chip with AI functions including edge and consumer devices. The exact endpoint is debatable; the trajectory is not.
The clearest demand signal in hard numbers comes from Nvidia. Its Q4 fiscal 2025 announcement (fiscal years run February to January) reported record quarterly revenue of $39.3 billion, up 78% year over year, with the Data Center segment alone hitting a record $35.6 billion — up 93% and over 90% of total revenue. Full-year fiscal 2025 Data Center revenue was $115.2 billion, up 142%. That same announcement named Nvidia a key technology partner for the $500 billion Stargate AI infrastructure project, and noted that more than 75% of the world’s TOP500 supercomputers run on Nvidia technology.
The pace has not slowed. Revenue hit $46.7 billion in Q2 fiscal 2026 (Data Center $41.1 billion, up 56% year over year), with the new Blackwell chips supply-constrained by production capacity — a capacity problem, not a demand problem. By Q3 fiscal 2026, revenue reached $57.0 billion (Data Center $51.2 billion, up 66%), with the CEO telling investors that cloud GPUs were “sold out.” On the latest reported tape, Nvidia’s Data Center revenue stands at $75.2 billion, up 92% year over year — while AMD’s Data Center business grew 107% to $6.7 billion and Broadcom’s AI semiconductor revenue surged 143% to $10.8 billion with a $30 billion backlog of custom silicon sold out through fiscal 2027. For context, the Semiconductor Industry Association puts the entire global semiconductor market at $791.7 billion in 2025, projected near $1 trillion in 2026 — meaning the AI-chip segment is now one of the largest single drivers of the whole industry.
The demand behind the demand is hyperscaler capital expenditure. In 2025, Amazon, Alphabet, Microsoft, and Meta together spent roughly $380 billion on capital projects, with about 75% of that AI-related — up from a combined total of about $24 billion in 2015, a 13- to 16-fold increase in a decade. CreditSights projects 2026 Big Five spending at roughly $602 billion, and Goldman Sachs puts total hyperscaler capex for 2025–2027 at $1.15 trillion.
| Year | Big tech capex | What changed |
|---|---|---|
| 2015 | ~$24B | Pre-AI baseline |
| 2024 | ~$256B | AI build-out begins in earnest |
| 2025 | ~$380B+ | ~75% AI-related (Amazon ~$125B, Alphabet ~$91–93B, Microsoft ~$80–95B, Meta ~$66–72B) |
| 2026 | ~$602B | CreditSights projection |
Where does that money go? Industry analysis breaks AI infrastructure spend into roughly $180 billion for GPUs and accelerators, $120 billion for data-center construction, $50 billion for networking, $40 billion for memory (HBM and DDR5), $25 billion for cooling, and $20 billion for power. Those are analyst estimates, but the ordering is the point: the chips themselves are the largest single line item. And the growth engine has shifted from training to inference. Training a frontier model is a one-time cost; answering billions of user queries a day is a permanent one. Every query needs accelerator time, which is why cost-per-token has become the metric vendors now optimize for.
Export controls and the fight over supply
AI chips are strategically important for another reason: almost nobody can make them. Taiwan produces more than 90% of the world’s most advanced logic chips through TSMC, and roughly 60% of all semiconductors come from Taiwan — the US imports 92% of its advanced logic chips from the island. The concentration extends to memory: SK hynix holds about 62% of the high-bandwidth memory (HBM) market, with Samsung and Micron making up most of the rest. In effect, four companies control the pricing power over the entire AI-chip stack.
Governments have noticed. The US CHIPS and Science Act of 2022 mobilized $280 billion in total ($53 billion in direct subsidies plus $24 billion in tax incentives) to rebuild domestic production. TSMC’s Arizona Fab 1 is operational on 4nm-class processes, with a second fab targeting 3nm in 2027 — always roughly one generation behind Taiwan, per the island’s export-control rules. TSMC received a $6.6 billion CHIPS grant plus $5 billion in loans; the company’s broader US investment commitments are reported at $165–265 billion, though those figures vary by report. Dependency has been reduced, not eliminated.
The most aggressive intervention has been US export controls on AI chips sold to China:
- October 2022: the Bureau of Industry and Security bans sales of Nvidia’s A100 and H100 to China. Nvidia responds with downgraded A800 and H800 versions.
- October 2023: BIS tightens the rules and closes the downgraded-chip loophole.
- 2024: Nvidia launches the China-compliant H20; in December, HBM exports are controlled and 140 entities are added to restriction lists.
- April 2025: the H20 also becomes license-required; Nvidia books a roughly $4.5 billion charge.
- July–August 2025: a reversal — the H20 is allowed again, with Nvidia and AMD paying the US government 15% of their China revenue.
- January 2026: the newer H200 moves to case-by-case review, with a 25% tariff and a 50% volume cap.
The net effect is contested, but analysts at Bernstein project Nvidia’s share of China’s AI-chip market falling from roughly 40% in 2025 to about 8% by end-2026, with Huawei’s share rising toward 50% — suggesting the controls accelerated Chinese indigenization as much as they slowed it. Either way, the policy arc proves the premise: AI chips are now treated as strategic assets on par with oil or uranium, with AI accelerators restricted by the same kind of export machinery once reserved for weapons technology.
Power bills, custom silicon, and what comes next
There is one more reason AI chips keep getting more important: they are hungry. AI accelerators consume roughly 2–4 times more energy than traditional CPUs, and in hyperscale AI data centers the compute hardware draws about 75% of facility power, versus roughly 60% for standard servers. The International Energy Agency estimates global data centers consumed 415 terawatt-hours in 2024 — about 1.5% of world electricity — and projects that will reach roughly 945 TWh by 2030, about 3% of world electricity, growing around 15% per year. Accelerated servers with GPUs drive about half of that net increase, and AI’s share of data-center electricity is expected to climb from roughly 25% in 2024 to about 49% in 2030. Goldman Sachs puts data-center power demand growth at about 50% between 2023 and 2027, reaching roughly 84 gigawatts. The hyperscalers’ response tells the story: Microsoft is spending about $1.6 billion to restart Three Mile Island (online 2028), Google has partnered with Kairos Power on modular reactors, and Meta has signed 6.6 gigawatts of nuclear agreements. AI demand is quite literally reviving nuclear power.
That power bill is also driving a rebellion. Hyperscalers are designing their own custom AI chips for inference workloads, aiming at 2–3 times better performance-per-dollar than buying Nvidia GPUs — cutting both the “NVIDIA tax” and the electricity cost. It is the biggest structural threat to Nvidia’s margins even as Nvidia keeps its grip on frontier training. Broadcom, which builds much of this custom silicon, reported $8.4 billion in AI semiconductor revenue in Q1 fiscal 2026, up 106% year over year, with a $30 billion backlog meaning its custom-silicon capacity is effectively sold out through fiscal 2027. Google’s v6 Trillium TPU is shipping (about 1.6 million units expected in 2026), its v7 Ironwood offers 4.6 petaFLOPS per chip with a 9,216-chip “Superpod,” and the Google–Broadcom partnership runs through 2031. AWS is shipping Trainium3 early in 2026 on 3nm processes, claiming 40–65% lower cost than comparable Blackwell setups. Meta’s MTIA v2 (“Artemis”) is deployed at scale for ads ranking and inference, with a v3 in planning for Llama-model inference.
Nvidia is answering with cadence. The company now launches a new architecture every year, and Vera Rubin — in production in 2026 on TSMC’s 3nm process with 288 GB of HBM4 memory at 22 TB/s of bandwidth, roughly 2.8x its predecessor, and about 50 petaFLOPS of inference performance per GPU — is paired with a new Vera CPU. Rubin Ultra (2027, four chips in a package using advanced chiplet packaging) and Feynman (2028) are already on the roadmap, all compatible with the same software stack that locks customers in.
The bottom line is that AI chip demand is structural, not cyclical. Chips get faster, data centers get bigger, the chips become obsolete — a roughly five-year replacement cycle is baked into the economics, according to analysts including IBM’s CEO — and the replacement orders are larger each time. When the world’s most valuable companies are building power plants to run their processors, those processors have stopped being components. They are the commodity the AI era runs on.
Top comments (0)