Three private companies – Anthropic, OpenAI, and SpaceX – are heading for public markets with combined valuations that exceed the total value of all US venture-backed exits over the past twenty-five years. The figure, reported by market analysts, nails an unprecedented phenomenon: never before have so few assets concentrated so much capital awaiting an IPO.

It’s not about the numbers. It’s the clearest symptom of a tech economy abandoning startup fragmentation to embrace an infrastructure oligopoly. Anthropic and OpenAI dominate the Large Language Model arena, while SpaceX leads in orbital launches and communications. Yet the common thread is their ambition to become the “operating system” of entire industries – from enterprise software to defense – through proprietary platforms requiring colossal hardware investments.

For those dealing with on-premise LLM deployment or local inference, these valuations should sound a warning. Capital concentration downstream translates into control over the GPU supply chain, compute nodes, and training pipelines. When three players absorb the lion’s share of the sector’s financial resources, the best chips – from NVIDIA H100s to future custom accelerators – end up in their data centers long before reaching the enterprise market. For an organization evaluating bare metal autonomy, the risk is being stuck in line for VRAM, facing extended delivery times and operating costs inflated by the giants’ demand.

There is, however, a flip side. The impending listing pushes the three companies to maximize recurring revenue, potentially accelerating the release of “on-premise” versions of their models to win over financial, healthcare, and government clients that never let data leave their own data center. Anthropic and OpenAI have already started offering self-hosted configurations for strategic customers, while SpaceX, through Starlink, enables private networks for military bases and remote platforms. The need to generate margins, combined with regulatory pressure on data sovereignty, could turn self-hosting from an exception into an official distribution channel, albeit at costs still prohibitive for most enterprises.

The likely losers are intermediate platform providers and startups that until now built services by licensing these models: once the LLM producer sells the on-premise instance directly, the middleman doing cloud hosting loses its foothold. At the same time, system integrators specialized in private AI infrastructure may get a new boost, because running a model with hundreds of billions of parameters locally demands orchestration skills, low-latency networking, and thermal management expertise that go well beyond the classic server rack.

Structurally, the mega-IPO story marks the end of a cycle where innovation was funded by hundreds of small, distributed VC bets. Now the market is wagering on a few “national champions” of AI, with valuations already pricing in platform dominance. For the on-premise ecosystem, the game will be played on the ridge between dependence on a handful of hardware roadmaps dictated by large funds and the chance to carve out sovereignty niches that no listed multinational can ever saturate.