The software movie is replaying in AI. And quickly.
Open-weight models have closed the capability gap with the closed frontier to single digits, from about thirteen points down to six on the Artificial Analysis Intelligence Index in roughly a year. Inference running on those open models now costs 6x to 60x less per token than the closed labs charge for the same job. We’ve watched this exact story before. Linux ate proprietary Unix. Postgres beat Oracle on every developer survey that matters. Kubernetes became the only way anyone runs containers. Next.js became the default way anyone ships a React app. Cheap and open keep winning. It just takes long enough each time that everyone forgets to expect it the next round.
Existing Used to Be the Whole Moat
Starting a technology company in the 2000s meant server racks, a real IT capital budget, and an enterprise sales team walking customers through an eighteen-month contract cycle. The product wasn’t really the moat. Being able to afford to exist was the moat. Building anything real took money and that kept the field small.
AWS launched in 2007, cloud infrastructure went mainstream, and servers turned into a line item instead of a capital project. Engineering didn’t get any easier, though. If anything it got scarcer. There were entire stretches where the consensus in Silicon Valley was that there simply weren’t enough software engineers. The ability to hire software engineers quickly became the bottleneck.
Open Source Started Giving the Answers Away
What actually broke that scarcity was open source, specifically the open core version of it. The hardest infrastructure problems, the ones a handful of companies had already solved internally at enormous cost, started showing up as software anyone could pull down for free. Top tier teams at companies like Confluent, HashiCorp, Databricks, Sentry, Vercel, Prefect, and JetBrains built and then gave away the secrets of high quality building. Each one took a problem that used to require its own engineering team and turned it into a dependency you add in an afternoon.
These companies didn’t stop making money, to be clear. Confluent, HashiCorp, Vercel, and Databricks are each worth billions. What changed is that the underlying capability stopped being the scarce, defensible thing. What used to be a unique engineering achievement became a commodity, and the company that built it made its money on top of the commodity instead of by hoarding it.
The Labs Thought They’d Broken the Pattern
AI labs looked at that whole arc and bet it wouldn’t apply to them. The pitch was that foundation models were different: capital intensive enough, talent constrained enough, and far enough ahead on compute that no open alternative could realistically catch up. A few labs went so far as to suggest they’d be the last startups anyone would ever need to build, since every other company would just be a thin wrapper on top of their API. Investors believed it enough to write checks in the billions on that premise alone.
But Open Source Always Finds a Way
Open-weight models are doing to the foundation model layer exactly what Kafka did to real-time data infrastructure and what Terraform did to cloud provisioning: taking what was a guarded, capital-intensive advantage and handing it to anyone willing to download the weights.
Flo Crivello runs Lindy, a 25-person company building AI automation. In June he moved all of Lindy’s traffic off Anthropic’s Claude and onto DeepSeek, expecting to save millions. Performance didn’t drop. On Lindy’s core use cases, it went up. He was quoted as saying: “It’s a matter of survival for the business. That’s all it is.”
Checkr fine-tuned an open Llama model for its hardest background-check cases and cut its monthly bill from around $12,000 to around $800, with responses landing 30x faster. Bridgewater fine-tuned an open Qwen model on its own financial data and reported it beat GPT-5 and Claude Opus on document triage, at something like a thirteenth of the compute cost, without a single record of client data ever leaving its own infrastructure. For a company in finance, healthcare, or anything else touching regulated data, that’s the difference between landing customers or being held up on security review.
We’re Not Betting on the Model
This dynamic is an existential matter to us at Supercruise. We’re not betting on DeepSeek, or Qwen, or whichever model wins the next benchmark cycle (Gemini just beat out OpenAI and Anthropic on their recent release, for instance). US intelligence agencies flagged in September that several Chinese labs had leaned hard on distilling US frontier models to help close that gap, which is a real governance question for anyone building on top of them, and a subject for another post. We’re not investing in the labs, and we don’t need our portfolio companies betting their infrastructure on any single one of them either. We’re investing in founders knowing that what founders build with whichever model is cheapest and capable this quarter and may switch next. The model is the commodity now, the same way Kafka and Terraform became commodities (though Confluent and HashiCorp thrived for reasons we can discuss). How people use - or harness - these models is where the valuable companies get built.
A Different Kind of Bootstrapped
VCs have spent two decades subsidizing the parts of a startup that used to be unavoidably expensive: the infrastructure, the engineering headcount, the eighteen months of building before there’s real revenue to fund it. That subsidy was, for a long time, the only way most founders could buy themselves the runway to exist. Take the servers, the reinvented wheels, and a meaningful share of the engineering headcount out of the cost structure, and a lot of founders don’t need to buy that runway from a VC anymore. We’re already watching founders skip the traditional venture round entirely and get further, faster, on a fraction of the capital.
This isn’t the bootstrapping most people picture: the lifestyle business that pays a founder a comfortable living and never grows past a small team. Scale is still the goal. What’s different is that the path there no longer has to run exclusively through endless rounds of funding.
There’s Always A Catch
None of this is free. It just moves where the cost shows up. Lower barriers to entry mean more competition, and not only from other startups. Incumbents get access to the same cheap infrastructure everyone else does, which means the old moat of “we built the hard part first” stops working for them too. Novel product bets, new business models, categories nobody’s named yet matter more now. And a clean cap table means more flexibility to react in real time. When the barrier to building drops this low, keeping your preference stack simple is one of the few structural advantages still fully in a founder’s own control.