This article is Part 1 of a three-part series on AI's real competitive barriers: where they're being built, who benefits, and what founders and business leaders should do about it.
“The model is not the moat” is Silicon Valley’s new “don’t be evil”: a mantra everyone repeats. The logic is that as AI models, the neural networks powering ChatGPT and its rivals, become increasingly commoditized and powerful, and as open-source alternatives mature, long-term competitive advantage can’t lie in the models themselves. Today’s breakthrough will become tomorrow’s baseline. The frontier keeps moving. What you built last year is what everyone can access for free next year. Venture capitalist Marc Andreessen calls it “a race to the bottom.” Therefore, the thinking goes, the real value must be elsewhere: in applications, in data, in user relationships.
That’s a soothing narrative. It implies that AI will be an open, competitive market where anyone can win. A Google memo from 2023 crystallized this sentiment: “We have no moat. And neither does OpenAI.” But follow the capital and a different picture emerges. In late February, OpenAI closed a $110 billion funding round, the largest private raise in history, at a post-money valuation of $840 billion, led by Amazon ($50 billion), Nvidia ($30 billion), and SoftBank ($30 billion). Weeks earlier, Anthropic closed a $30 billion round at a post-money valuation of $380 billion, the second-largest private raise on record. Meanwhile, OpenAI has committed over $350 billion in cloud contracts with Azure and AWS, Amazon has poured $58 billion into OpenAI and Anthropic combined, and the big labs are signing exclusive data deals with everyone from Reddit to the Associated Press to Shutterstock. Nvidia’s best chips are pre-sold through 2027. If AI models are really becoming interchangeable, low-margin commodities as the model-moat adage implies, someone forgot to tell the people writing the checks.
As a venture investor and startup operator in healthcare, I've been drawn to AI by the same instinct that draws any entrepreneur to a wave they can feel but not yet fully see: the need to understand what's coming before it arrives. I've spent the past year sitting with innovators, investors, execs, and healthcare providers who are all asking the same question: what is actually going on?
Here’s what I’ve concluded: AI is poised to eat the world, and healthcare and pharma are on the menu. But the “no moat” narrative is a misdirection. The barriers are absolutely being built, just not in the places everyone insists are commoditizing. The model isn’t the moat. But compute is. Data is. Distribution is. Power is. Regulatory capture is. Once you control those inputs, a flywheel kicks in. More power, training capacity, and proprietary data produce better models, which attract more usage and capital, which then buy you even more of the underlying inputs. The companies saying “don’t worry, this will all be open and competitive” are the same ones quietly locking up every input that matters.
The Compute Chokehold
“Compute” is the processing power to train and run AI models. It has become the scarcest resource in tech, making training frontier models a capital expenditure and logistics endeavor. And the barrier to entry is access to Nvidia’s latest data‑center GPUs (H100, H200, Blackwell, and beyond). OpenAI’s $250 billion to Azure is an advanced booking dressed up as a cloud contract. They’re locking in priority access to Microsoft’s biggest clusters for years to come. It reads like a keep‑out sign.
Amazon is hedging both sides. $8 billion into Anthropic bought deep integration with its custom Trainium silicon. $50 billion into OpenAI, announced last week, bought AWS the role of exclusive third-party cloud distributor for OpenAI’s enterprise platform. That’s $58 billion to ensure Amazon sits at the infrastructure table no matter which model wins. Good luck competing with that if you’re a Series B startup trying to rent capacity on the spot market.
And Nvidia? Jensen Huang has basically pre-allocated his entire supply chain to an inner circle of frontier labs (OpenAI, Anthropic) and hyperscalers (the cloud giants: Amazon, Microsoft, Google). Everyone else is fighting for scraps. Startups are waiting quarters, not weeks, for cluster access. In AI time, that’s an eon.
The whole game has inverted. It used to be: build a great model, then figure out compute. Now it’s: secure compute, or don’t bother showing up.
And compute is only half the equation. Training at this scale requires gigawatts of power. Data centers already consume 4% of US electricity, with some projections putting that figure above 12% by 2030. The big players are scrambling to lock up the power generation to run all their compute. Microsoft has signed a deal to restart Three Mile Island. Amazon’s buying data center sites near nuclear plants. Sam Altman, OpenAI’s CEO, chairs fusion startup Helion Energy and is a major backer of advanced fission startup Oklo, reflecting his bet that both fusion and fission will be needed to power AI-scale compute. With multi-year lead times and heavy capital intensity, power infrastructure is yet another barrier to entry. Once you’ve cornered compute and power, the next logical step is to corner what flows through those machines: the data.
Who Owns the Corpus?
Here’s something that deserves more attention. The data wars are already over. The incumbents won. OpenAI has licensing deals with Shutterstock, the AP, Reddit, News Corp, and others. With lawsuits piling up against AI companies over training data scraped without permission, these licenses provide legal cover while also locking in privileged access to high-quality training data that open-source projects can’t easily replicate.
The open-source models everyone’s excited about have a data problem. You can replicate an architecture. You can even match performance on benchmarks. But you can’t replicate a curated corpus (the training data) of licensed images, news archives, and human-curated Q&A. That data doesn’t exist on the open internet. It’s locked up.
The conversation fixates on parameter counts and benchmark scores. The data-licensing deals get covered as news, but their strategic implications as durable competitive advantages get far less attention. If you’re building a foundation model on Common Crawl while OpenAI is training on the entire AP archive back to 1985, you’re not in the same fight. You’re in a different weight class entirely.
OpenAI is paying News Corp $250M+ over five years for access to its archives. Shutterstock has deals with OpenAI, Google, Meta, Amazon, and Apple. Compute is the fuel and data is the oxygen. The frontier labs are cornering both markets while the debate centers on openness and access.
So the next time someone tells you the model is not the moat, ask a different question: if that’s true, why are the biggest players in tech spending hundreds of billions to lock up compute capacity and training data? The model may not be the moat — but the inputs to the model already are. And by the time most people realize it, the gates will be closed.
Coming up in Part 2:
In Part 2 of this three-part series on the AI landscape, I’ll explore how tech giants are using distribution to lock in users at planetary scale, why regulation is warfare in disguise, and ask whether open-source is really a viable alternative.
If you found this article valuable, share it with a founder, investor or executive navigating the AI landscape.
Thanks for reading Backed By Reality.
Subscribe for free to receive new posts.
Interesting read.