AI private valuations and $1tn of mark-ups with unit economics still unproven

• Private AI valuations have exploded while revenue remains concentrated and infra costs are heavy.

• I would prefer picks and shovels or disciplined co-invest alongside apps with clear payback, not 100x ARR at seed.

What happened

On 17 October 2025, reporting highlighted that ten lossmaking AI start-ups including OpenAI, Anthropic and xAI have added almost $1tn in value over the past 12 months, with US VCs deploying $161bn into AI year to date. Early-stage AI apps with about $5m ARR are seeking $500m plus valuations, well above 2021 peaks.

Context & data

• AI’s share of global VC reached an unusually high proportion of total funding in 2025, reflecting rapid capital rotation into model and application layers, per recent Venture Monitor commentary from PitchBook in April 2025.

• OpenAI’s annualised revenue has been reported at roughly $13bn three years after ChatGPT’s launch, indicating exceptional top-line growth relative to peers as of mid October 2025.

• Enterprise data and AI platform spend is scaling, with leading data infrastructure vendors signalling multi-billion dollar revenue run-rates through 2025 which suggests CIO budgets are shifting toward AI-adjacent software and tooling.

• VC deployment into AI is on course to exceed $200bn in 2025, outstripping prior software investment cycles and raising the bar for future cash generation.

My take

This feels like reflexivity at work. Hyperscaler and chip capex enable bigger models. Bigger demos trigger enterprise pilots. Rising pilots justify higher secondary marks and bigger infra commitments. The top assets may earn their marks, but the average app at 100x ARR will not. The path to software-like cash generation depends on two things. First, normalised gross margin after inference costs needs to move up as models and serving get cheaper. Second, distribution must compress CAC and shorten payback. Most start-ups will only have one of those.

Where would I lean in? I like infra-adjacent software that sells into existing CIO budgets. Think data platforms, governance and orchestration with 80% plus gross margin, net revenue retention above 120%, and sub-twelve-month payback once embedded. I also like applications that ride a privileged channel. That might be native inside the data stack or via strong partners. I want evidence of durable cohorts and contribution profit after compute, not just usage growth. I would avoid horizontal copilots with commodity models unless there is proprietary data, a compliance moat, or captive distribution.

Risks & watch-list

• Compute and power constraints that inflate COGS or cap growth

• Concentration risk on a single cloud or foundation model provider

• Regulatory drag in the UK and EU across safety, data and cloud competition

• Funding fragility if secondary marks reverse or credit tightens