Three Things from All-In's Liquidity Summit That Matter to a Small Firm

The All-In podcast’s Liquidity Summit (Napa Valley, 31 May–3 June) brought together the bosses of the biggest AI and security firms with the investors who fund them. Three of their keynotes matter a lot to UK small businesses — not because you will ever meet these people, but because the decisions they are making this quarter shape the tools you already pay for.

The cyber CEO’s six-week miracle

Palo Alto Networks’ CEO Nikesh Arora told the summit his team pointed a new AI model called Mythos at their own software. “In six weeks we found what would have taken us five to seven years,” he said. The cost: “low millions” of dollars in compute. He added that, with the model’s “ultra mode” — a setting that keeps it reasoning until it lands an answer — his team could chain weaknesses together into full attack paths.

5–7 yrsof manual security-audit work done in 6 weeks by an AI model, per Palo Alto Networks CEO Nikesh Arora at the All-In Summit — for a “low millions” spend in compute.

This is not a demo. It is a real cyber-defence firm running AI against its own production code. Arora also described a 250-person marketing team whose output is now “90% consistent” and a 5,000-person sales floor that “acts almost consistently” with customers. The capability ceiling jumped in the last twelve months, and it is no longer restricted to Silicon Valley.

For a small firm, the practical read is direct: an afternoon’s work with an AI code reviewer (most paid tiers now ship one) can spot the obvious holes in a website, an order form, or an internal script before someone else does. If you would rather not send the code off-site at all, our piece on running a capable local model on a 24GB GPU is the place to start.

The $100bn+ compute bill

OpenAI’s CFO Sarah Friar confirmed the company raised $122bn privately in March and will spend north of $100bn on the chips, data centres and power behind the models. Her rule of thumb: 1 gigawatt of compute capacity is worth roughly $10bn a year of revenue. “In 2026, we still won’t have enough compute,” she said, and on a 1-gigawatt site in Michigan the company is promising to “pay for our infrastructure and our power” so ratepayers are not left with the bill.

For a UK small firm this translates to a mixed picture. The price you pay per token is still falling (good), but the supply of high-end model capacity is being rationed (bad). Expect occasional throttling, regional slowdowns, and feature rollouts that land in the US months before they reach UK tenants. We have covered how to budget for usage-based, agentic AI pricing and why the $16–$20 a month tier is now the de facto standard — but assume that tier will keep shifting as the labs renegotiate their own deals.

The same compute squeeze is why a sovereign UK option matters; our piece on the £500M Sovereign AI Unit explains where Isambard-AI and the AI Growth Zones fit in.

The IPO wave that’s coming

Thomas Laffont of Coatue told the summit a roughly $4tn wave of AI IPOs is in the pipeline — SpaceX, OpenAI, Anthropic and others. His data point that matters most to small firms: companies at unicorn stage (over $1bn valuation) have only an 8% chance of becoming a $10bn firm, but companies already at $100bn-plus have a 31% chance of ten-xing again. Translation: the rich get richer; the smaller players get bought, merged, or shut.

That has direct consequences for the AI tool on your credit card today. The vendor you pay in June may not exist on its own in two years. Arora told the summit his own firm cut a 20-seat SaaS product to three users, pointed Claude at the data via Slack, and “reduced our bill by 90%.” That is the template: keep your data portable, keep your prompts documented, and do not build critical workflows around a single vendor.

What to do this week

• Run an AI code review on your own systems today. Ask your paid model (ChatGPT, Claude, Gemini — all do it) to find security issues in your website, booking page, or any customer-facing script. Free if you are on a paid tier; otherwise see our free tiers roundup.
• Audit your AI subscriptions for vendor risk. Which of your tools come from a company that could plausibly be acquired in eighteen months? Make sure you can export your data and prompts if they vanish.
• Decide what stays on your own box. Anything sensitive — customer data, internal docs, code — is worth running locally. Our LM Studio vs Ollama 2026 comparison helps you pick the runtime.
• Track usage, not seats. With compute rationed, per-task pricing will keep shifting. Our guide to tracking AI agent spend is a one-afternoon project.

The summit speakers were not talking to you. But the decisions they announced this week will set the price, reliability and longevity of the AI tools you already rent. Plan accordingly.