TL;DR
- Major AI companies from OpenAI to SpaceX are all moving toward designing their own chips, because they want to control both cost and supply chain instead of relying solely on Nvidia’s long queue
- The impact on Nvidia right now is more of a “warning” than a “shake-up” — the chips each company designs are mostly used to supplement specialized workloads in-house, not sold directly in the open market
- For readers, look at this trend as a signal that the AI industry is entering an era where hardware and software can no longer be separated — whoever controls their own chip design gains a long-term edge in performance-per-cost
(Note: this article discusses the industry trend at a high level and does not cite specific sales figures or market share numbers.)
When chips become a strategic weapon, not just hardware
Honestly, this isn’t just about “who’s faster” — it’s about “who controls more of their own destiny.”
Normally, the GPU is a shared resource everyone buys from a single supplier. When demand spikes, prices spike along with it, and you still have to wait in line for delivery.
Having your own chip is like a factory having its own supply chain — no need to depend on anyone, and you can design it to fit your own workload precisely.
That’s why OpenAI, Google, Amazon, and even SpaceX have all moved toward building their own silicon — it’s about seizing the single most important bottleneck in the entire AI industry.
The starting point that made everyone rethink doing their own chips
Rewind to when the AI boom first ignited: every company rushed to book Nvidia H100s, with queues stretching months, even years.
Some companies had already paid and were still waiting for delivery, because manufacturing capacity couldn’t keep up with the surging global demand.
As supply ran short, GPU rental prices on the cloud spiked right along with it. The cost of training large models became a massive, uncontrollable expense, because everything hinged on a single supplier.
That’s the point where AI executives started asking: can we really keep betting our entire future on someone else’s chips?
Who stands where in this equation
Google has been playing this game the longest — its TPU chips have trained and run Gemini in-house for several generations, making it the company least dependent on Nvidia in the group.
Amazon and Microsoft also have their own chip families (Trainium/Inferentia on Amazon’s side, Maia on Microsoft’s side), used primarily for workloads within their own cloud, cutting long-term GPU rental costs.
Meta has also built inference chips for its own ranking and ads systems. OpenAI is the newest entrant, partnering with Broadcom to co-design a chip manufactured through TSMC — they’re not building their own fab, just doing the design in-house.
On Elon Musk’s side, Tesla and SpaceX use Dojo/AI4 chips focused specifically on inference for cars and autopilot systems — different from the group above, which focuses on data centers.
Nvidia itself still stands at the center of the market, because its GPUs are the most flexible — anyone can use them without being locked into a single company’s ecosystem.
How different is the first generation from the latest?
Comparing each company’s first-generation chips to their latest ones makes it clear why everyone keeps investing further. Most first-gen chips still had weaknesses in connecting large numbers of chips together (interconnect) and incomplete software support, so they still had to rely on supplementary GPUs.
The latest generation is designed to fix those weaknesses directly — focusing on linking chips into larger clusters and tuning the software stack more closely to their own model-training workloads. The direction is to reduce reliance on Nvidia for tasks they can handle themselves, not to compete by selling chips on the open market.
Note: concrete performance/energy/cost figures for each generation are not yet confirmed in this article. Anyone wanting real numbers will need to wait for independent benchmarks comparing the companies.
| Factor | First-generation custom chip | Latest-generation custom chip |
|---|---|---|
| Large-scale chip interconnect | Still limited | Improved to scale further |
| Software support | Still relies on supplementary GPUs | Tuned specifically for own workload |
| Primary goal | Experimental/partial cost reduction | Long-term reduction of Nvidia dependence |
When custom chips show up in real-world work
Picture a model-training team that used to wait weeks for GPU access. Once the company has its own in-house chips, the queue shortens, and work that used to be blocked everywhere starts flowing more smoothly.
Finance teams feel it just as much. Cloud bills that used to spike with rented GPU usage become more predictable once hardware is controlled in-house — a number the CFO can price products around with more confidence.
But small startups are still living in a different world — they don’t have the capital to design their own chips, so they still depend on Nvidia as before, just with longer waits, since big companies are now buying less (having their own chips), which may spread demand out somewhat.
In the end, the picture is this: whoever has the capital to build their own chips gains an edge in long-term cost control, while whoever doesn’t is still playing the same game with Nvidia.
Head-to-head with Nvidia and rivals in the same arena
Once you’re actually in the ring, each player picks a different strength. Nvidia wins on ecosystem and software (CUDA), which has dominated the market for years, while the custom chips from Google, AWS, and Microsoft were built specifically to solve the cost and lock-in problem.
The clearest difference is availability — whoever has their own chip doesn’t have to wait in line for GPUs like other companies. Lock-in, meanwhile, flips around — instead of being locked into Nvidia, you end up locked into the cloud provider that owns the chip.
| Factor | Nvidia GPU (H100/B200) | Custom chip (TPU/Trainium/Maia) |
|---|---|---|
| Ecosystem/software | Most complete (CUDA) | Tied to the owning cloud |
| Long-term cost | Pay according to Nvidia's margin | Control cost in-house |
| Availability | Must wait in queue | Available immediately in-house |
| Lock-in | Tied to CUDA | Tied to cloud provider |
Pros and cons of moving to build your own chips
Now that we’ve seen the overview, let’s look at whether investing in custom chips is really worth it — because not every company that tries this survives it.
The clear upside is controlling long-term cost yourself, not waiting in Nvidia’s purchase queue, and designing the chip to fit the job exactly — for example, a chip made only for model training doesn’t need unnecessary extra features.
But the downside is just as heavy: chip design requires enormous investment from R&D all the way through fabrication, and there’s a risk of falling behind technologically if you can’t keep pace with Nvidia’s annual updates.
Most importantly, you have to compete with world-class R&D teams that have accumulated decades of experience — having money alone doesn’t mean you can pull it off.
Pros
- +Control long-term cost yourself, without paying according to Nvidia's margin
- +Reduce reliance on someone else's supply chain — available immediately within your own group
- +Fully customize the chip to fit your own workload
Cons
- −Requires enormous investment from design all the way through fabrication
- −Risk of falling behind technologically if you can't keep up with market updates
- −Must compete with world-class R&D teams with decades of accumulated experience
The cost nobody talks about in the headlines
Headlines love to say “Company X has now designed its own chip,” but the real cost is far bigger than that.
First is the cost of chip architects — extremely scarce engineers fought over across industries. Sky-high salaries aren’t even enough; you may wait years just to have a team ready for production.
Second is the fabrication queue at places like TSMC, booked out for years in advance. Whoever arrives late still gets in line behind Apple and Nvidia.
Third is geopolitical risk — the chip manufacturing supply chain is concentrated in just a handful of countries. When tariff issues or export restrictions hit, a whole year’s plan can collapse overnight.
Finally, the thing most often forgotten is the software/driver stack that must be developed in parallel with the hardware. No matter how good the chip is, if the software stack isn’t ready, it can’t actually be used. That’s exactly why only truly large companies dare to play this game.
So how much longer can Nvidia keep controlling the game?
In the short term, Nvidia still has almost no competitor for large-scale training, because the CUDA ecosystem — built over a decade — isn’t something that can be copied overnight. But what has changed is that the “custom chips” from OpenAI, SpaceX, and Google aren’t here to replace Nvidia across the board — they’re here to replace only specific specialized inference workloads where a generic GPU is more expensive than necessary.
What’s really worth watching over the next 2-3 years is “decentralization,” not “dethroning the champion.” Major AI companies will gradually reduce their reliance on Nvidia workload by workload, gradually thinning Nvidia’s margins.
For investors, watch the share of Nvidia’s revenue that comes from its 4-5 biggest customers. For developers, watch whether the software stacks of these custom chips open up enough to attract users outside the company — that’s the deciding factor for whether they remain just internal tools or become genuine competitors.