AI and specialised chips: Who wins and loses as demand surges?

NVIDIA became more valuable than the GDP of countries like Saudi Arabia and the Netherlands, on the back of chips, even as many of its biggest customers are still running at a loss.

Key takeaways

NVIDIA controls more than 70% of the AI chip market, a concentration of infrastructure power the tech industry has never quite seen before.
Global AI chip spending is projected to surpass $163 billion by 2026, nearly tripling from $53 billion in 2023.
Legacy giants like Intel, once untouchable in semiconductor dominance, are visibly losing ground in the AI era.
U.S. export controls are quietly splitting the world into two AI tiers: those who can access cutting-edge chips, and those who can’t.
For AI startups today, GPU access is as much a fundraising conversation as it is a technical checkbox.

Every major technological shift has had its defining infrastructure moment. The internet had fiber optic cables and server farms. Mobile had the baseband chip. AI has the GPU, and it has been a shockwave.

What started as a hardware component built to render video game graphics has quietly become the most strategically important piece of technology on the planet. The AI boom has since moved from being just a software story to becoming a supply chain story, a geopolitics story, and increasingly, a story about who holds the keys to the next decade of economic power.

This article will explain why chip demand exploded the way it did, who’s winning, who’s getting left behind, and what the next few years could look like for an industry that’s moving faster than most governments and corporations can keep up with.

How AI created an unprecedented chip demand curve

For most of computing history, the CPU was king. It was fast, versatile, and built to handle tasks one after another in rapid sequence, the digital equivalent of a brilliant multitasker.

Then AI came along and broke that model entirely.

Training a large language model or running computer vision at scale isn’t a sequential task. It’s thousands, sometimes millions, of mathematical operations happening simultaneously. Matrix multiplications, tensor calculations, massive parallel workloads. CPUs weren’t built for that kind of punishment. GPUs were.

Related Story: 10 things you need to know about Gemini enterprise

Originally designed to push pixels in video games, GPUs turned out to be almost perfectly architected for the math that powers modern AI. That accidental alignment between gaming hardware and machine-learning workloads ignited the demand curve nobody fully anticipated.

When ChatGPT crossed 100 million users in two months, faster than any consumer application in history, the world suddenly needed a lot more of what only a handful of companies knew how to make.

The winners: who is gaining a structural advantage

NVIDIA: the infrastructure monopoly

Depending on the segment, NVIDIA controls somewhere between 70% and 95% of the AI chip market, making it the critical infrastructure layer for the entire AI industry. The company posted $215.9 billion in revenue in fiscal year 2026 (ending January 2026), a staggering 65% increase from the $130.9 billion reported in the previous fiscal year.

NVIDIA’s proprietary CUDA platform has built a 15-year moat of deep-seated expertise and infrastructure. For AI labs and engineers, switching hardware is more than just a procurement decision; it’s a massive, unwanted migration project that threatens years of established workflow and code.

TSMC (Taiwan Semiconductor Manufacturing Company)

If NVIDIA is the face of the AI chip boom, TSMC is the factory floor. The Taiwan-based foundry manufactures more than 90% of the world’s most advanced chips, including those made by NVIDIA, AMD, Apple, and Google.

Experts say this is arguably the single largest systemic risk in the global tech economy right now. TSMC’s concentration of advanced manufacturing capability means that any disruption, for whatever reason, affects the entire AI build-out simultaneously.

Hyperscalers

Google, Amazon, Microsoft, and Meta have all clearly read the NVIDIA dependency problem and are doing something about it. Each has developed or is actively developing custom silicon:

Google’s TPUs (Tensor Processing Units).
Amazon’s Trainium & Inferentia.
Microsoft’s Maia.
Meta’s MTIA (Meta Training and Inference Accelerator).

Custom chips optimized for specific workloads can dramatically cut inference costs at hyperscaler scale. The catch is that none of this pays off in the short term. We’re looking at a three-to-five-year horizon before these investments meaningfully shift the balance. Until then, they’re still buying NVIDIA, just with a plan to buy less of it eventually.

AMD

AMD is still pulling its weight. Its MI300X chip is a credible alternative to NVIDIA’s H100, and more importantly, it offers a lower-cost option at a time when GPU access is rationed like a scarce commodity. For cost-sensitive AI deployments, that’s a real opening.

The honest limitation is the ecosystem. CUDA’s depth and maturity are things AMD’s ROCm platform hasn’t matched yet. But AMD is gaining ground, and in a market this large, even a meaningful slice of the challenger position translates to billions in revenue.

The losers: who is falling behind

Intel

The company that defined the processor era for decades posted a $1.6 billion loss in Q2 2024 and announced layoffs of 15,000 employees, roughly 15% of its global workforce.

While NVIDIA was quietly building CUDA and deepening its GPU ecosystem, Intel stayed committed to CPU dominance, a position that AI workloads have systematically devalued. Being the best at sequential processing doesn’t help much when AI needs thousands of things at once.

Mid-tier cloud providers

The cloud infrastructure market is consolidating in real time. Mid-tier providers have seen their collective market share fall, a compression driven by the superior spending of the “Big Three” hyperscalers (AWS, Microsoft Azure, and Google Cloud), who now collectively command roughly 63% of the global market.

When hyperscalers are deploying tens of billions annually to secure chip supply and build custom silicon, smaller providers are competing with a hand tied behind their back. Customers chasing cutting-edge AI capabilities increasingly have fewer reasons to look beyond the top three.

China’s AI sector

China’s AI ambitions are real and well-funded, but U.S. export controls have introduced a ceiling that money alone can’t break through. Restricted from accessing NVIDIA’s most advanced chips, Chinese AI developers are largely working with hardware that performs at a fraction of the H100’s capability.

That gap compounds over time: slower training runs, higher compute costs, and a widening distance from the frontier models being built on unrestricted hardware. Domestic alternatives like Huawei’s Ascend chips are improving, but catching up to TSMC-manufactured, CUDA-optimized silicon is a multi-year problem, not a near-term fix.

AI startups

For early-stage AI startups, GPU access is now a fundraising variable. In addition to asking about your model architecture or your go-to-market, investors are also asking what your compute situation looks like, whether you have a cloud partnership, and how much of your runway disappears into GPU bills every month.

The startups winning this game tend to have one of two things:

A hyperscaler partnership that gives them preferential access to compute.
Enough investor backing to secure GPU allocation ahead of demand.

Everyone else is building in a resource-constrained environment, hoping that efficiency will close the gap. Sometimes it does. Often, it doesn’t.

What comes next?

The inference opportunity

Most of the chip conversation has centered on training, the GPU-intensive process of building foundation models. But inference, the act of actually running those models at scale, is where the next hardware competition is taking shape.

Inference chips have different optimization requirements, and that distinction opens the door for challengers that the training market largely kept out.

New names worth watching

Groq, Cerebras, and Qualcomm are all positioning for a slice of the inference market, each with architectures purpose-built for speed and efficiency at deployment scale. None of them is NVIDIA yet. But in a market shifting this fast, “not NVIDIA yet” can still mean a multi-billion dollar business.

FAQs

Why are specialized chips critical for AI?

AI workloads run on parallel computation, processing millions of operations simultaneously. Standard CPUs aren’t built for that. Specialized chips like GPUs are why they became indispensable almost overnight.

Can anyone actually challenge NVIDIA?

Yes, and in many ways, they already are. AMD is closing ground, and hyperscalers are building their own silicon. The pressure is real, but NVIDIA’s CUDA ecosystem gives it a lock-in that won’t loosen quickly.

Why is AI so expensive to build?

Compute costs dominate everything. Training a single model runs between $50 million and $200 million.

Conclusion

What’s happening in AI chips is a structural realignment of where power lives in the global tech industry.

NVIDIA, TSMC, and the hyperscalers are cementing advantages that compound with time. Intel, mid-tier cloud providers, and chip-restricted markets are navigating a hole that gets harder to climb out of the longer it deepens.

Citations

Disclaimer!

This publication, review, or article (“Content”) is based on our independent evaluation and is subjective, reflecting our opinions, which may differ from others’ perspectives or experiences. We do not guarantee the accuracy or completeness of the Content and disclaim responsibility for any errors or omissions it may contain.

Related Story: Nigeria eyes passage of AI law to lead regulation push in Africa

The information provided is not investment advice and should not be treated as such, as products or services may change after publication. By engaging with our Content, you acknowledge its subjective nature and agree not to hold us liable for any losses or damages arising from your reliance on the information provided.

Always conduct your research and consult professionals where necessary.