The frontier sat quiet today — no new model card, no closed mega-round — and when the labs go quiet, the action moves to the two places access is actually decided: the consumer tier and the finance layer. Both moved today. Google confirmed it will pull Gemini Flash and Pro out of the free tier tomorrow, October 9, leaving free accounts Flash-Lite-only, in the same week OpenAI started pushing GPT-6 out to free users. And the money on compute split in public: Lambda is raising up to $4 billion in a final round at a $14.5 billion pre-money on the strength of Anthropic’s giant lease, while a few time zones away an Nvidia-backed neocloud IPO just cratered for lack of orders. The week that the models went quiet, the market re-priced access.
The free tier was a funnel, and Google decided the funnel’s work is done
What happened. Starting tomorrow, October 9, Gemini free accounts lose the models they could actually lean on — Gemini 3.6 Flash and the limited 3.1 Pro access free users get today — and drop to Flash-Lite only. It is not just the free tier: AI Plus at $5 a month also loses Pro, keeping Flash-Lite and Flash, so any real Pro access now starts at the $20 AI Pro tier or the $100 Ultra. Google put the whole move on one dated schedule, which is the useful part: it is a plan you can migrate against, not a rumor.
Why it matters. Read it as an engineer: the free tier was never free, it was a funnel — priced to acquire, budgeted as a marketing line item, and carrying genuinely good models so the habit formed before the paywall came up. Google just decided the funnel has converted what it will convert, and took the expensive models off the budget. If your ad-hoc bulk lane — doc-heavy summarization, quick coding questions, structured extraction — quietly depended on Gemini’s free or near-free tier, that lane compresses to the smallest model tomorrow, and the re-budget is now on a date you can schedule against. The tell is the contrast in the same week: OpenAI is moving the opposite direction, rolling GPT-6 to all tiers with the cheaper Luna powering free accounts as of today. Two labs, one week, two opposite calls on what consumer access is for. Google is treating the free tier as an exhausted funnel; OpenAI is treating it as a still-open battlefield. Yesterday’s column flagged October 9 as the first real signal on whether free consumer access is a racket or a funnel — tomorrow lands the first dated answer.
Source: digitaltrends.com
The neoclouds are raising like frontier labs while the public market cools
What happened. Lambda, the AI neocloud renting capacity to the frontier, is raising up to $4 billion in a “final round” before a planned 2027 IPO — around a $14.5 billion pre-money with Coatue and Blackstone leading — on a backlog the WSJ puts at $15 billion growing toward $50 billion, much of it traced to Anthropic’s roughly $35 billion compute deal. The counter-signal landed in the same window, from Sydney: Bloomberg reports Firmus Grid’s planned $5.5 billion listing is “shrouded in uncertainty” after its A$11 marketed price failed to attract adequate support — a company with $51 million of revenue, a 912-megawatt pipeline of which only 46MW is actually built, and a valuation that nearly tripled in two months.
Why it matters. This is price discovery happening in public, and the two curves just split. Private rounds price the compute boom — Lambda raising $4 billion pre-IPO on a backlog its anchor customer wrote — while the first real neocloud listing in front of the public order book balked. For anyone renting capacity, the vendor’s financing story is not background noise: a cratered IPO or a gapped round changes a neocloud’s buying power, and that changes your rack price and your vendor’s odds of existing in two years. The Firmus numbers do the telling on their own — $51 million of revenue carrying a valuation built to nearly triple in two months on 46MW of built megawatts. The paper is ahead of the steel, and the public market just said so, on the same compute-finance layer this column has tracked since Lambda wrote Anthropic’s September lease.
Source: bloomberg.com, wsj.com
The Rest
- Nous put a funnel on the open agent — “Hermes for Businesses” is live with the $1.5B valuation and full backer list confirmed (Robot Ventures leading; Nvidia, M12 and Y Combinator among them), and the number worth tracking is the conversion curve: roughly $36M ARR in mid-September toward an expected >$100M by year-end on 24M clones, about 2.5% of global token usage — yesterday’s balance-sheet take just acquired a sales motion, so watch the open license terms as the seats scale. techcrunch.com
- DeepSeek’s ≥$12B is still not closed — third straight day in the column — every wire recirculates the same Bloomberg thread (close to securing at least ¥80B, Tencent and CATL at the head, early-2027 IPO prep), and the honest caveat is still CNBC’s: the amount “remains fluid” and could change before talks conclude; nothing has closed since the September finalization. cnbc.com
- Manus raised $500M+ in its first round since Beijing killed the Meta deal — Boyu and IDG leading with Tencent following on, in what’s being called China’s largest single AI application-layer round; a clean read on where Chinese capital flows while every frontier raise sits in talks. cnbc.com
- OpenAI gave the free tier a real model the same week Google took one away — GPT-6 with Intelligent UI began rolling to paid tiers October 7 and reaches Free and Go today, powered there by the cheaper Luna; the clearest possible contrast to Google’s narrowing move above. openai.com
- The Pentagon ordered CENTCOM ready to resume strikes on Iran — per Axios, the order could be exercised before the November 3 US midterms or Israel’s October 27 election; the military thread this region watches stays live into the vote. timesofisrael.com
- San Jose’s pause pressure ramps two days before its own gate — residents and the Sierra Club’s Loma Prieta chapter are pushing for an immediate data-center moratorium while the city drafts “uniform standards” ahead of the October 15 session on its 30-plus-project pipeline, the council-chamber constraint closing in on the South Bay heartland. sanjosespotlight.com
What I’m watching
- October 9: the Gemini downgrade lands — the first dated evidence on whether free consumer access was a racket or a funnel, and OpenAI’s free-tier GPT-6 rollout makes the comparison test cleaner than it was yesterday.
- October 16: RTX Spark goes on sale — whether the $2,599 floor holds, and the first independent shots at the ~120B-on-a-laptop claim.
- October 27: Mistral’s Large 4 weight drop, license included — now that the RTX Spark tier gives open weights a local home to run on.
- Close-watch: DeepSeek’s ≥$12B, OpenAI’s $30B at $1.4T, Lambda’s $4B, and Nscale’s still-blank S-1 range — plus whether Firmus’s cratered bookbuild is the start of a pattern in neocloud listings.