What if the Genesis Mission were a huge mistake?

A few days ago it dawned on me that my daughter has spent months using a Chinese artificial intelligence model to prepare her Master's assignments, and that in all that time she hasn't felt the slightest need to pay for a subscription to any of the industry's big players. When I asked her whether she didn't miss Claude's power, she gave me that look, a mixture of condescension and affection that children reserve for their parents' absurd questions, and said: "Dad, the tokens run out faster and faster, and for what I need, this is more than enough."

"Good enough disruption"

That phrase has haunted me ever since. Because it sums up with chilling precision the concept of "good enough disruption" that Clayton Christensen formulated in The Innovator's Dilemma: market leaders tend to overserve their customers with features most of them don't need, opening the door to an inferior but much cheaper competitor that captures the base of the pyramid... and that, from there, climbs up until it devours their business.

Since this article is going to be long, I ask you to hold on to that idea while I splash around in a seemingly unrelated puddle.

But... what is the Genesis Mission?

In November 2025, President Trump signed an executive order launching the "Genesis Mission": a national effort led by the Department of Energy to apply frontier artificial intelligence to scientific research, with rhetoric that explicitly compared it to the Manhattan Project. The aim is to connect the DOE's 17 national laboratories to share their supercomputers and their decades of experimental data in fusion, fission, materials science and biotechnology, and to apply to them the most powerful models from OpenAI, Anthropic and Nvidia, among others. To date, 51 organizations have joined the effort and $320 million has been committed in initial investment, with an additional $150 million just to curate and structure the existing scientific data.

Quite a wake-up call, adding to the Stargate Project previously endorsed by the US government: a consortium of OpenAI, SoftBank, Oracle and Nvidia that aims to invest $500 billion by 2029 in training and inference for next-generation models.

Is the world turning upside down?

It is the sum of these two initiatives that makes me wonder, more and more, whether the premise might be inverted.

Because the Genesis Mission assumes that the scarce asset is the frontier model, and that competitive advantage comes from putting it to work on data that already exists. But what if the truly scarce asset were the data, and the model were interchangeable? If a Chinese model that costs a hundredth of the Western one is "good enough" to process those same experimental data, all this investment in frontier capacity becomes dispensable. Or extremely dangerous, if we consider the inertia that so much investment will generate in assets whose obsolescence turns them into a ticking time bomb. To paraphrase that line from Bill Clinton's campaign, "It's the data, stupid, not the infrastructure."

And the symptoms that the market is moving in that direction keep piling up. DeepSeek has just launched its V4 Flash model at $0.14 per million input tokens, between 30 and 100 times cheaper than its OpenAI equivalent. Models derived from Alibaba's Qwen family already exceed 100,000 variants on Hugging Face, more than any Western model. And a partner at Andreessen Horowitz estimates that 80% of US startups use Chinese models as the foundation for developing their products.

Read that again, please: 80% of the startups in the country that has just launched the Genesis Mission are building on the models of its geopolitical rival.

A (to say the least) debatable strategy

But if the cost issue seems worrying to me, the strategy is no less risky.

While the United States invests sums that would make the Marshall Plan blush (the Stargate Project promised $500 billion in AI infrastructure, and planned investment among the five big hyperscalers for 2026 is around $700 billion), China has opted for a radically different path. Instead of competing on raw power, it has opened up its base models and made them available to anyone who wants to build on them. According to a report by the US Congress's Economic and Security Review Commission, this approach generates a double feedback loop: Chinese labs refine each other's models without costly investments in pre-training, and companies around the world adapt them to specific niches, generating usage data that feeds the improvement of the original models. The Chinese aren't competing to have the most powerful engine; they're building the gas station where everyone fills up.

We've been here before

Does this ring a bell? It does for me: the flat rates of the telecom operators. At the turn of the century, the telcos financed the highways on which Google, Netflix and Amazon built their empires. The uncomfortable question is whether the Genesis Mission (and the entire Western AI strategy) might be repeating the same logic: investing massively in infrastructure so that whoever controls the data and the vertical application captures the real value.

And there is one more fact that invites deeper reflection. The marginal returns of frontier models are starting to falter. The recent launch of Claude Opus 4.7 was met with a barrage of complaints from developers: measurable improvements in programming tasks, yes, but accompanied by regressions in web research, a tokenizer that makes usage between 12% and 18% more expensive, and false positives in safety filters that block legitimate tasks. It is the classic Christensen pattern: marginal improvements in cutting-edge features that most people don't need, while the overall experience deteriorates.

That said, it's worth being honest about the limits of this thesis. Christensen got the iPhone wrong: he predicted its failure because cheap phones were "good enough". And it can't be ruled out that frontier models will unlock qualitative capabilities (drug discovery, materials design, scientific reasoning) that justify their cost for a niche willing to pay whatever it takes. Genesis could turn out to be right if those capabilities prove decisive. Moreover, the DOE's classified data won't be run through open-source Chinese models for reasons of national security, which gives Western models an advantage that is not one of quality but of trust.

The final irony

But there is a certain irony in all this: the very architects of the Genesis Mission have devoted $150 million to curating and structuring the DOE's data before they can feed it to any model at all. It is an implicit acknowledgment that the valuable asset is not the model but the data. If we're right about all this, Genesis is aiming the power of supercomputers and frontier models at the wrong problem. And the true mission of a new "Manhattan Project" should be to build the best datasets in the world, not the biggest computing infrastructure. Because the engine costs a little less with every quarter that goes by. But the fuel is irreplaceable. And it is, probably, the only asset in which you are, without a doubt, ahead of your competitor.

My daughter, meanwhile, oblivious to the strange things her father thinks and writes, keeps using a Chinese model that costs her nothing and works perfectly for her.

Perhaps we should start asking ourselves whether we are looking at the finger when the moon is somewhere else.