There is a piece of hardware that every major tech company, every government, and every serious AI lab on the planet desperately wants more of. It is made by one company. That company's market capitalisation has oscillated between $2 trillion and $3.5 trillion. Its CEO wears a leather jacket to every major conference and is treated like a rockstar.
The hardware is Nvidia's Blackwell GPU. The CEO is Jensen Huang. And the story of how a semiconductor company became the infrastructure layer of the 21st century's most important technology is one of the most remarkable in business history.
What Is Blackwell?
Blackwell is Nvidia's GPU architecture released in 2024–2025, succeeding Hopper (the architecture behind the H100 chip that dominated AI training from 2022 onward). It is named after David Harold Blackwell, a pioneering statistician and mathematician.
The flagship chip in the Blackwell family is the GB200, which Nvidia claims delivers up to 30x better performance than the H100 on large language model inference workloads. The GB200 NVL72 — a rack-scale system containing 72 Blackwell GPUs connected with Nvidia's NVLink technology — has become the most coveted piece of enterprise hardware since the mainframe era.
To put the numbers in context:
- The H100, released in 2022, was already the fastest AI chip available and sold for $25,000–$40,000 per unit
- The GB200 NVL72 rack system costs approximately $3 million per rack
- Lead times for delivery stretched to 12+ months at peak demand in 2025
- Microsoft, Google, Meta, Amazon, Oracle, and virtually every major cloud provider pre-ordered at a scale that strained Nvidia's supply chain
Why GPUs? Why Nvidia?
To understand Nvidia's dominance, you need to understand why AI training requires GPUs in the first place.
Training a large language model involves running billions of mathematical operations — specifically matrix multiplications — simultaneously. CPUs (the processors in standard computers) are designed for sequential tasks. They are fast at running one thing after another.
GPUs were originally designed for rendering graphics — a task that requires doing thousands of simple calculations in parallel (one for each pixel). It turns out that this parallel processing architecture is almost perfectly suited to the mathematics of neural networks.
Nvidia has been making GPUs since 1999. But the critical advantage they built wasn't just the hardware — it was CUDA.
The CUDA Moat
CUDA (Compute Unified Device Architecture) is Nvidia's software platform for programming GPUs. Launched in 2006, CUDA gave researchers and developers a way to use Nvidia GPUs for general computation, not just graphics.
When deep learning took off in the early 2010s — starting with the landmark 2012 AlexNet paper that ran on Nvidia GPUs — the AI research community was already fluent in CUDA. Every framework, every library, every research paper was written to run on CUDA.
This created a moat that competitors have spent over a decade failing to cross:
- AMD makes excellent GPUs but lacks CUDA's ecosystem depth
- Intel has tried multiple times to enter AI chips with limited success
- Google built its own TPUs (Tensor Processing Units) — excellent for Google's own workloads but not available externally
- Startups like Cerebras, Groq, and SambaNova have built interesting architectures but remain niche
The result: if you want to train a frontier AI model today, you almost certainly do it on Nvidia hardware running CUDA. The switching cost is enormous — not because of the hardware, but because of the software ecosystem built around it.
The Blackwell Supply Chain Crisis
When Blackwell launched, demand immediately outstripped supply in ways that bordered on absurd.
The chips are manufactured by TSMC (Taiwan Semiconductor Manufacturing Company) using its most advanced 4NP process. TSMC has a near-monopoly on cutting-edge chip fabrication — there is no other company on the planet currently capable of manufacturing chips at this level of complexity and scale.
The supply chain for a Blackwell GPU involves:
- TSMC fabricating the dies in Taiwan
- TSMC's CoWoS packaging connecting the GPU die to HBM (High Bandwidth Memory) stacks
- SK Hynix or Samsung supplying the HBM3e memory
- Foxconn or Quanta assembling the final server systems
- Nvidia managing the entire orchestration
Every step is a potential bottleneck. In 2025, HBM supply was the tightest constraint. The advanced packaging capacity at TSMC was another. The result was a market where $3 million racks were being resold at a premium, and smaller companies simply couldn't access the hardware they needed.
This is why the phrase "compute is the new oil" has become cliché but remains accurate — whoever controls the most advanced AI compute infrastructure controls the pace of AI development.
Who Is Buying Blackwell (And How Much)?
The scale of investment in Blackwell infrastructure is staggering. Here's a snapshot of publicly announced commitments:
| Company | Announced Spend | Notes |
|---|---|---|
| Microsoft | $80B (FY2025) | Largest single-year AI infrastructure spend in history |
| Meta | $60–65B (2025) | Building 2GW+ of AI data centres |
| $75B (2025) | Expanding TPU and GPU capacity | |
| Amazon (AWS) | $100B+ (multi-year) | Trainium + Blackwell mix |
| Oracle | $40B+ | Aggressive expansion for AI cloud |
| xAI (Elon Musk) | ~$10B | Memphis "Colossus" cluster |
These are not speculative bets — they are infrastructure commitments based on the belief that AI demand will continue to grow faster than supply for the foreseeable future.
The Geopolitical Dimension
The AI chip race is not just a technology story. It has become a central front in US-China geopolitical competition.
The US government, beginning under the Biden administration and continuing under Trump, has imposed increasingly strict export controls on advanced AI chips to China. Nvidia was initially prohibited from exporting its H100 to China, then the A100, then downgraded versions created specifically for the Chinese market (the H800 and A800).
In response:
- Huawei has accelerated development of its Ascend AI chip series
- Biren Technology and other Chinese chipmakers are racing to close the gap
- Chinese tech giants (Baidu, Alibaba, ByteDance) are stockpiling whatever chips they can access
- The Chinese government has made semiconductor self-sufficiency a national strategic priority
The gap remains significant — most estimates put Chinese AI chips 2–4 generations behind Nvidia's latest. But the trajectory is one of gradual closing, not permanent divergence.
For Nvidia, China's exclusion from its customer base represents a loss of what was once ~25% of its data centre revenue. For the US, the question is whether export controls slow China's AI development enough to matter — or whether they simply accelerate Chinese domestic chip investment.
What Comes After Blackwell?
Nvidia moves fast. The roadmap is already public:
- Blackwell Ultra (B300) — enhanced version of Blackwell, released in 2025
- Rubin — next architecture, expected 2026, named after astronomer Vera Rubin
- Rubin Ultra — expected 2027
Jensen Huang has promised a new architecture every year. The cadence has accelerated from the historical two-year cycle to an annual one, driven by AI demand.
The underlying bet Nvidia is making: the demand for AI compute is not a temporary spike. It is a permanent structural shift — every company, every government, every major institution will eventually need to run AI models at scale. And for the foreseeable future, that means Nvidia hardware.
The Risk: Can Anyone Catch Up?
Nvidia's dominance looks unassailable today. But history suggests that technology moats, even deep ones, can be eroded.
The credible threats:
Custom silicon from hyperscalers: Google's TPUs, Amazon's Trainium, Meta's MTIA, and Microsoft's Maia are all becoming more capable. These chips are designed for specific workloads and optimised for each company's own models. They will never be sold externally, but they reduce hyperscaler dependence on Nvidia for their own inference workloads.
AMD MI300 series: AMD has made genuine progress. The MI300X has found real customers, particularly for inference. It remains behind Blackwell, but it's a credible alternative for some workloads — and the ecosystem gap is narrowing.
Software disruption: The most underappreciated risk. If someone builds a widely-adopted AI framework that abstracts away the hardware layer, CUDA's moat becomes less relevant. OpenAI's Triton compiler is one attempt at this. It hasn't succeeded yet, but the direction is clear.
Physics: Even Nvidia is not immune to the limits of semiconductor physics. As chips approach atomic scale, the traditional improvement curves slow. Nvidia is betting that 3D chip stacking, new memory architectures, and system-level integration can sustain performance gains even as per-transistor improvements slow.
Why This Matters Beyond Tech
The AI chip story is really a story about who controls the foundational infrastructure of AI — and therefore, in some meaningful sense, who shapes the development of the technology.
Right now, that power is concentrated:
- One company dominates AI chip supply (Nvidia)
- One company dominates advanced chip fabrication (TSMC)
- A handful of hyperscalers own most of the compute infrastructure
This concentration has implications for AI safety (who can audit and control frontier models?), economic competition (can startups access the compute they need?), and geopolitics (will democracies maintain a lead in AI capability?).
Jensen Huang often says that the more you buy, the more you save — because using AI accelerates the work that justifies buying more AI infrastructure. It is a flywheel that, once spinning, is very hard to slow down.
The Blackwell era is not just a product cycle. It is a glimpse at what the AI economy looks like at scale — and a reminder that in every technological revolution, the people who build the picks and shovels often end up wealthier than the people hunting for gold.
Want to understand the broader picture? Read our articles on how AI agents actually work and vibe coding — how AI is changing software development.