DC Decoded
Six Months Later: The AI Industry Is Acting Like One Big Startup
The layers still exist. What changed is how quickly Nvidia, Anthropic, OpenAI and the hyperscalers are willing to change the plan while executing it.
Ross is probably my favourite character in Friends, and one of my favourite scenes is the famous “Pivot! Pivot! Pivot!” episode.
After my recent stint in the US, I actually found myself in almost the same pose while moving furniture. It felt like déjà vu.
But the word pivot also reminds me of something very different.
During my MBA, whenever we discussed building a business or a startup, one lesson came up again and again: be ready to pivot.
Pivoting is not necessarily a sign that the original idea failed.
When you start a business, you begin with a view of the customer, the problem, the product and the business model. Then the market teaches you things you did not know.
You learn.
You adjust.
Sometimes you change the product. Sometimes the customer. Sometimes how you make money. Sometimes you realise that what you thought was your biggest advantage is not enough.
That flexibility is part of building a startup.
Which brings me back to AI.
AI has already become a huge industry. Some of the largest and most experienced companies in the world are involved. The amount of money being invested is extraordinary.
Yet when I look at the industry today, I still see something that behaves remarkably like a startup.
Not one company.
The entire industry.
Companies are still figuring out what they should own, what they should buy, what they should build themselves and what they cannot afford to depend on someone else for.
Sometimes the move is about growth.
Sometimes it is protection.
Sometimes it is insurance in case the original strategy turns out to be wrong.
And sometimes the market simply moves faster than anyone expected.
Six months ago, I wrote that AI was not one race.
I saw four different theories of how to win.
OpenAI was trying to own almost everything — from energy and data centers to chips, models and the final product.
Anthropic seemed more focused on building the best model and using infrastructure built by others.
Google, Amazon, Microsoft and Meta were building the platforms and infrastructure everyone would need.
And Nvidia owned the chips underneath almost all of it.
Six months later, I still think those layers exist.
What I missed was how quickly everyone inside them would be willing to change the plan while executing it.
Pivot. Pivot. Pivot.
Only this time, the couch is worth billions of dollars.
Nvidia moved beyond the chip before it needed to
On September 3, Nvidia agreed to buy Hugging Face for $12.93 billion.
Hugging Face is not a chip company.
It is one of the biggest places where AI developers find, share and build with models. More than 18 million developers, researchers and creators use it, and more than 200,000 companies use the platform.
Nvidia says Hugging Face will remain open. Developers will still be able to use other models, frameworks, clouds and computing platforms.
That is what makes the move interesting.
Nvidia does not need everyone on Hugging Face to use only Nvidia hardware.
It needs Nvidia to remain important somewhere in the process of building and running AI.
Nvidia is still far ahead in AI chips today. It did not need Hugging Face to sell more GPUs this quarter.
That is exactly why I find the timing interesting.
Google has TPUs.
Amazon has Trainium.
Meta is developing its own chips.
Broadcom is growing rapidly in custom AI silicon.
At some point, today's extraordinary GPU growth will slow.
Nvidia seems to be using its position today to make sure it still matters when that happens.
Reuters called the Hugging Face deal something close to an insurance policy. I think that description fits.
That is a pivot made from strength rather than desperation.
And those are usually the pivots worth watching.
NVIDIA to Acquire Hugging Face — NVIDIA, September 3, 2026 ↗
Nvidia wraps itself in $12.9 billion of insurance — Reuters Breakingviews, September 3, 2026 ↗
Anthropic moved from using compute to securing it
Six months ago, I saw Anthropic as almost the opposite of OpenAI.
Build a great model.
Let Amazon and Google spend the money building chips and data centers.
Use their infrastructure rather than trying to own everything yourself.
That description does not quite hold anymore.
Anthropic still does not appear interested in building the entire infrastructure underneath Claude.
But it clearly no longer wants infrastructure availability to decide how quickly Claude can grow.
In April it committed more than $100 billion over ten years to AWS technologies, securing up to 5 gigawatts of new capacity.
It already uses more than one million Amazon Trainium2 chips.
Separately, it expanded its partnership with Google and Broadcom for multiple gigawatts of next-generation TPU capacity starting in 2027.
And more recently Reuters reported that Anthropic is forming its own custom-silicon team.
What makes this particularly interesting is that Anthropic had previously been cautious about making enormous infrastructure commitments. Then demand grew much faster than expected.
The strategy changed with it.
A great model is not enough if you cannot get enough computing power to run it.
Anthropic has not suddenly become an infrastructure company.
It pivoted from betting mainly on the model to also betting on never being the reason Claude cannot serve the demand when it arrives.
Anthropic and Amazon expand collaboration for up to 5GW of new compute — Anthropic, April 20, 2026 ↗
Anthropic expands partnership with Google and Broadcom — Anthropic, April 6, 2026 ↗
OpenAI moved the competition from answers to work
OpenAI is interesting for a different reason.
Six months ago, I often found Claude better for some coding and long-form work.
Today, for me, that gap feels much smaller.
But that is not the bigger change.
The bigger change is that I increasingly spend less time thinking about the model itself.
With ChatGPT Work, I can start with an outcome rather than a prompt.
I can bring together the context around a piece of work, let it research across different sources and files, analyse what matters, move between tools and keep refining the same piece of work until there is something usable at the end.
That is a very different experience from asking one model a question and comparing its answer with another.
As a user, the competition starts moving from:
Which model answers this better?
to:
Which system understands enough of what I am trying to do to actually carry more of it through?
That is much more interesting to me.
Because once AI becomes part of how I research, analyse, write, work across information and produce something at the end, the underlying model is still important.
But it is no longer the whole experience.
OpenAI has not stopped competing on model quality.
It has widened what it is competing for.
It is no longer only trying to win the prompt.
It is trying to own more of the work that happens before and after it.
That is another pivot.
ChatGPT is now a partner for your most ambitious work — OpenAI, July 9, 2026 ↗
The hyperscalers may be making the biggest move of all
Google, Amazon and Microsoft were already infrastructure companies in one sense.
They built cloud computing.
But AI appears to be pulling them much deeper into the physical side of that business.
They remain some of Nvidia's biggest customers.
At the same time, Google has TPUs.
Amazon has Trainium and Inferentia.
Microsoft has Maia.
Meta is developing four new generations of its MTIA chips in two years, while openly saying it will continue to use outside silicon where that makes sense.
That last point is interesting.
Meta says AI models are changing faster than normal chip-development cycles. A chip can take years to reach production, and by then the workload it was designed for may already have changed.
Its answer is not to make one perfect long-term bet.
It is to shorten the cycle.
Build.
Learn.
Change.
Build again.
That sounds remarkably like startup thinking.
But the hyperscaler story is much bigger than chips.
These companies build data centers.
They secure power.
They buy Nvidia chips.
They design competing chips.
They rent computing capacity to other AI companies.
They build their own models.
And then they sell AI products on top of all of it.
Microsoft recently reorganised its financial reporting into two major groups.
One of them is literally called “Agents and Infra.”
I find that name surprisingly revealing.
Infrastructure underneath.
AI doing the work on top.
Increasingly, both are part of the same strategy.
At the scale these companies are now investing, AI is no longer only a software or technology race.
It is also becoming a race for chips, land, data centers, financing and power.
Four MTIA Chips in Two Years — Meta, March 11, 2026 ↗
Microsoft reveals Azure cloud sales in financial reporting shift — Reuters, September 2, 2026 ↗
The layers did not disappear. The strategies became less rigid.
I do not think Nvidia has stopped being a chip company.
Anthropic has not stopped believing that Claude has to be one of the best models.
OpenAI has not stopped competing on model quality.
Google, Amazon and Microsoft have not stopped being cloud companies.
Their original strengths still matter.
But nobody seems comfortable depending on only one of them.
Nvidia has chips, but is moving closer to developers and models.
Anthropic has Claude, but is securing infrastructure and exploring its own silicon.
OpenAI has ChatGPT and frontier models, but is moving deeper into chips, infrastructure and the actual work people want AI to do.
The hyperscalers have cloud infrastructure, but they are building chips, models, agents and AI products as well.
Everybody seems to be buying some insurance against being wrong.
Or opening another source of revenue.
Or protecting themselves from a supplier.
Or simply following where the market is taking them.
That may be the biggest thing I missed six months ago.
I thought I was looking at four different strategies.
I now think I am looking at an industry that is still learning what the winning strategies actually are.
Software can pivot faster than infrastructure
There is one big difference between this and a normal startup.
A startup can change its product, pricing or customer and write off a few months of work.
These companies are changing strategy after committing to physical things.
Chips.
Data centers.
Power contracts.
Factories.
Transmission infrastructure.
Billions of dollars of equipment.
And those things do not pivot nearly as quickly as software does.
That is what makes this phase particularly interesting to me.
They are learning like startups.
But they are making infrastructure-sized bets while they learn.
A wrong turn can stay on the balance sheet for years.
Where I land
I still do not think there will be one winner in AI.
If anything, I am more convinced of that than I was six months ago.
But I no longer see four groups running four separate races with fixed strategies.
I see an industry that is still finding its shape.
Nvidia is learning.
Anthropic is learning.
OpenAI is learning.
Google, Amazon, Microsoft and Meta are learning.
And they are changing direction while the race is already running.
AI may already be one of the biggest technology industries the world has seen.
It may have some of the largest and most experienced companies in the world behind it.
But strategically, it still behaves remarkably like a startup.
Pivot. Pivot. Pivot.