The GPU made Nvidia the most valuable company in history.

The chip that replaces it is already being built by Google, Amazon, Meta, Microsoft and OpenAI simultaneously.
Here is what that means for where the money actually flows next.
First, you need to understand why the GPU became king.
Go back to 2012.
A neural network called AlexNet obliterated the competition at a global image recognition contest using a single insight.
The parallel processing Nvidia built into GPUs to render lifelike graphics in video games is structurally identical to what training a neural network requires.
Thousands of smaller cores performing matrix multiplication simultaneously rather than a small number of powerful cores running sequential tasks.
One researcher took a GPU and hacked it to expose those parallel computation capabilities for deep learning.
That moment started a decade long run that turned Nvidia from a gaming company into the infrastructure layer of the entire AI economy.
Six million Blackwell GPUs shipped in the last year alone.
A single 72-GPU Blackwell server rack sells for approximately $3 million.
Nvidia is shipping one thousand of them every week.
For a brief moment in October, Nvidia became the first company in history to reach a $5 trillion valuation.
That is what one insight from 2012 compounded into.
But the GPU has a structural problem that nobody was talking about loudly until recently.
It is a Swiss Army knife.
Extraordinarily capable across a wide range of AI workloads, but not optimised for any single one of them.
In the early boom era of large language models, that flexibility was the point.
Training required massive general purpose parallel compute and the GPU delivered it better than anything else available.
But as models matured, the balance shifted.
Post-training techniques have made models increasingly capable.
Now the dominant workload is not training.
It is an inference.
Every time you open Claude, ChatGPT, Gemini, or any AI product and receive a response, that is inference.
Every Starbucks app transaction, every Salesforce workflow, every AI assistant running in your EarPods.
All inference.
And inference can run on less powerful chips programmed for more specific tasks.
That single shift in workload balance is what opened the door for the chip that is now being built to challenge Nvidia's dominance.
The ASIC is the chip that changes the map.
Application Specific Integrated Circuit.
Where a GPU is a Swiss Army knife, an ASIC is a single purpose tool.
Hardwired to do the exact mathematical operations for one type of job.
Faster at that job, more power efficient at that job, and significantly cheaper to operate at scale for that job than any general purpose GPU.
The trade-off is flexibility.
Once carved into silicon an ASIC cannot be reprogrammed for a different workload.
But for companies running inference at the scale of billions of daily requests, that trade-off is not a disadvantage.
It is precisely the point.
Google was first.
The TPU, Tensor Processing Unit, launched in 2015 and helped lead to the invention of the transformer architecture in 2017.
The transformer is the foundation of virtually every modern AI system running today.
Google's seventh generation chip, Ironwood, just launched alongside a deal to train Claude on up to one million TPUs.
Amazon built Trainium and Inferentia after acquiring an Israeli chip startup in 2015.
Anthropic is currently training its models on half a million Trainium2 chips inside an Amazon data centre in Indiana with no Nvidia GPUs in the building.
Trainium delivers between 30 and 40 percent better price performance than competing hardware vendors on AWS according to Amazon's own data.
Meta has its own training and inference accelerator.
Microsoft has its Maia chips targeting Azure data centres.
OpenAI is building custom ASICs in partnership with Broadcom starting in 2026.
Every major hyperscaler is building its own chip simultaneously.
Not as an experiment.
As a strategic infrastructure decision worth hundreds of billions of dollars in capital commitment.
Broadcom is the name most people outside the chip industry have not priced into their thesis yet.
Every major hyperscaler with an ASIC programme partners with at least one chip design company for the IP, the engineering expertise, and the networking infrastructure that connects the chips at scale.
Broadcom dominates that market.
Google's TPUs. Meta's training accelerator. Now OpenAI's custom ASICs.
Analysts tracking this space estimate Broadcom is winning 70 to 80 percent of the custom ASIC backend market.
That market is projected to grow at a mid double digit compound annual growth rate over the next five years.
The ASIC wave is accelerating faster than the GPU market.
Broadcom sits at the centre of almost all of it.
Then there is the edge layer that most people are not following yet.
As data centre AI matures the next battleground is on-device inference.
Your phone. Your car. Your laptop. Your wearables.
The Neural Processing Unit is the chip that powers AI locally without sending data back to a cloud server.
Privacy, speed, and cost efficiency all improve when inference runs on the device rather than in a data centre.
Qualcomm dominates NPUs for Android.
Apple's M-series chips include a dedicated neural engine for MacBooks.
The A-series chips in the latest iPhones have neural accelerators built in.
AMD and Intel are competing for NPUs in Windows laptops.
The dollars are concentrated in data centres today.
But the volume of chips required to put AI into every phone, car, robot, and wearable on Earth is an order of magnitude larger than the data centre market.
That transition has already started.
The geopolitical layer underneath all of this is the constraint nobody in the consumer narrative is talking about enough.
Almost every chip in this entire ecosystem, Nvidia Blackwell, Google TPU, Amazon Trainium, Apple A-series, is manufactured by one company.
Taiwan Semiconductor Manufacturing Company.
TSMC.
The concentration of advanced node semiconductor manufacturing in Taiwan is the single largest geopolitical chokepoint in the global AI race.
The CHIPS Act started the process of building TSMC fabs in Arizona.
Nvidia's Blackwell is now in full production at the Arizona facility.
Intel is manufacturing advanced node chips at a new Arizona fab.
But Apple's latest iPhone chip still requires TSMC's three nanometer process which is currently only available in Taiwan.
The reshoring of semiconductor manufacturing to the United States is happening but the timeline is measured in years not months.
And China is building its own parallel stack.
Huawei, ByteDance, and Alibaba are all developing custom ASICs under export controls that limit their access to the most advanced equipment and Nvidia's Blackwell chips.
The AI chip race is not just a technology competition.
It is a geopolitical infrastructure war being fought in silicon.
The country that secures the most advanced manufacturing capacity and the most reliable power supply for running it wins something far more valuable than a market.
Here is the frame that ties all of it together.
Nvidia earned its position.
Years of developer ecosystem investment, CUDA as a proprietary software moat, and a hardware roadmap that stayed ahead of every competitor built one of the most durable competitive advantages in technology history.
That advantage does not disappear overnight.
But the market is getting so large that it creates room for an entirely new layer of winners to emerge alongside Nvidia rather than simply replacing it.
The hyperscalers reducing their Nvidia dependence through custom ASICs.
Broadcom capturing the backend infrastructure of every major ASIC programme simultaneously.
Qualcomm and Apple owning the edge inference layer as AI moves onto every device.
TSMC remaining the irreplaceable manufacturing chokepoint regardless of which chip architecture wins.
And underneath all of it the power infrastructure required to run everything at scale becoming the constraint that determines who can actually build at the speed the AI race demands.
The GPU made Nvidia the most valuable company in history.
The companies who understood that 2012 moment before it became obvious never needed to explain their timing.
The same insight gap exists right now in the ASIC transition.
The people paying attention to the chip layer underneath the model race are already positioned.
The people who act on this tonight will understand why tomorrow.