Ali

Same night, two prices for AI software. Private: Cognition, maker of the Devin coding agent, is weighing offers at $70–80B, per Business Insider. It raised at $48B in September on almost $900M of run-rate revenue. That's roughly 53x then, 78–89x now. Public: ServiceNow grew subscription revenue 21%, posted a 98% renewal rate and guided a 36% free cash flow margin. Its Q4 bookings guide came in about half a point light. The stock fell 17% after hours. Both prices are the same bet: agents take work that used to be sold by the seat. My exit math on the private side: buy at $75B, want 3x, exit at a generous 20x revenue. That needs ~$11B of revenue, over 12x today's run-rate. Possible. Cognition's run-rate almost doubled between May and September. But the round only works if growth like that holds for years and margins stay software-like while the models underneath reprice. When a late-stage AI allocation reaches a Gulf desk through an SPV with fees on top, I ask one thing: does this still work if the next round is flat? If not, it's momentum, not underwriting.

The market sold Alphabet on cash burn. I'd read the Cloud line first. Google Cloud: $30.7B revenue, up 103%. Operating income $11.1B, up from $3.6B a year ago. Margin went from 24% to 36%. Of every extra dollar of Cloud revenue, about 48 cents became operating profit. Much of that profit is someone else's compute bill. Google says AI labs, capital markets firms, financial institutions and enterprises are driving the demand. Same night, Microsoft said Foundry revenue more than tripled. So when an AI startup tells me "model prices keep falling, our margins will expand", I ask who keeps the saving: them, or the cloud they run on. What I'd fund: products whose costs drop when they switch models or bring capacity in-house. What I'd pass on: a margin story that depends on a hyperscaler's pricing staying kind. For Gulf money, this quarter makes the case for owning compute, not just funding the apps on top of it.

Anthropic gave 50 professional analysts a Claude Code agent with a secret, coin-flip opinion on four judgment calls. They went with the agent 86% of the time. Chance is 50%. When they had already stated a view, 60% switched if the agent disagreed. Afterwards, 77% defended their picks with the agent's arguments. 15% mentioned the agent. Read that as an investor. When a memo starts as an agent draft, on the founder's side or ours, the numbers get checked. The judgment calls inside it often don't: which metric, which comps, which downside case. What I'm changing: • the partner writes the call down before opening the agent's memo • the agent has to make the best case for the other side, every time • in diligence, I ask founders which of their key numbers an agent picked What I'd fund: tooling that shows where a model's lean entered a decision. Gulf banks and funds rolling out agents will need that trail before their regulators ask for it.

26 companies have now told the OKX–ICE tokenized stock venue: not our shares. Microsoft, Alphabet, Goldman Sachs, Broadcom. Since Monday's notice, IBM, Oracle, Intel and IonQ too. Of the 62 names on the venue's first list, 23 have said no (CoinDesk's count). The rule makes no the easy answer. Under the SEC exemption, a company gets 30 days to object to a third-party token of its stock. Objecting costs nothing. Staying silent can't be taken back later. Same week, Janus Henderson launched JQQQ: a Nasdaq-100 fund whose shares are issued onchain, licensed from Nasdaq, for non-US professional investors, $100k minimum. That's the line I'd invest on: • a token of someone else's stock, made without them: the company can switch it off, and 26 just did • a fund or issuer putting its own shares onchain: nobody upstream to object In Gulf tokenisation pitches my first question is now who signed off. If the answer is "nobody had to", I pass.

An AI startup for wealth advisors just raised $70M at a reported $500M valuation. Two years old, about 20 customers. The number behind it isn't seats. It's assets on the platform: $50B to $700B this year, now adding ~$75B a month. Revenue up more than 25x (no figure given). That's the metric I'd underwrite in wealth tech. It grows when the customer's book grows, and once accounts open through you, you're hard to rip out. What I'd fund in the Gulf: • the boring middle: onboarding, KYC, suitability, custodian paperwork. Nevis says it cut account opening from weeks to days • built for DIFC and ADGM rules, Arabic and English, multiple custodians What I'd pass on: another meeting note-taker for advisors. Every CRM will ship that as a feature. DIFC alone has 592 wealth and asset managers and 1,408 family-related entities. Whoever builds the Nevis for them gets my meeting.

The number that matters in Broadcom's $63B Anthropic deal isn't 63. It's 72.5%. Banks and Blackstone fund an SPV that buys Broadcom chips and leases them to Anthropic. If the racks end up worth less than 72.5% of what's still owed on the lease, Broadcom covers the gap. Peak exposure: about $43B, running down to zero over five years of lease payments. The lease covers Anthropic's credit. The guarantee covers what nobody can price yet: what these chips are worth late in the lease. And it's not the first. On an earlier $35B Apollo/Blackstone deal, Broadcom's backstop was up to $29B. Moody's has already flagged the risk if these stack up. How I read it as an investor: • when the seller guarantees the floor, the seller is the real underwriter • in any compute or hardware pitch, I now ask who holds the residual, not just who signed the lease • when these SPVs get offered to Gulf allocators, the first question is which layer. The second lien sits under ~$43.5B of senior debt

Apollo's chief economist says AI has delivered "zero financial payoff" to corporate America outside tech. His charts: • revenue per worker: Mag 7 up to ~$780k. The other 493 in the S&P: ~$525k, still below its 2022 peak • profit margins: Mag 7 near 30%. The 493: ~10%, flat • no sign of AI in productivity growth yet That 493 is the customer list in most AI decks I see. If the buyer books tokens as a cost centre, the startup's revenue is the first line cut in a bad quarter. Pilots aren't proof. What I'd fund: AI priced on a number the CFO already tracks. Invoices collected, claims closed, tickets resolved. What I pass on: per-seat copilots with pilot revenue and no renewal data. In the Gulf, a lot of corporate AI buying is top-down right now. Year-two renewals will show who actually moved a margin.

Lenders didn't shut the door on AI. They shut it on GPUs with nobody contracted to pay for them. Same Nvidia chips, very different price of money: • SpaceX: $40B of GPU debt, backed by Starlink and launch revenue. Even so, its tranches were expected around 7.75–8.3% (Semafor) • CoreWeave: its 2031 bond briefly yielded over 16% this month (Reuters) • Nebius: a $3.5B Citi facility last week, pointing to $40B+ of long-term contracts with investment-grade customers Saudi's Humain runs the same playbook. Its CEO: "When you have bankable off-take deals, raising capital for me is not an issue." What I ask any compute pitch now: • who prepaid, and how much of the capex? • contract length vs the chip's useful life? • who pays for the idle racks in 2027? No signed offtake, no cheque.

1/7 I went through this week's MENA startup news. The money was small. The permissions were big. Biggest disclosed startup round in Arab News' weekly wrap: $3.6M. The licence news out of the UAE and Saudi mattered more than any of those cheques. What happened, and what I now ask every Gulf fintech pitch: