Jason Lemkin：S-1启示、裁员与AI模型路线图实操

What founders should do differently after Anthropic’s S-1, Microsoft’s buyout offer, Cloudflare’s 20% cut, and a Qwen model trying to reach a server in Beijing

Every Thursday, Harry, Rory and I go through the week’s news on 20VC x SaaStr. Here are the 10 things from this week’s episode I’d want a founder to act on.

1. Anthropic Went From Founding to IPO in 5 Years. SpaceX Took 24.

Anthropic was founded in 2021. It filed its S-1 on October 26 and its IPO, targeting roughly $2 trillion, is slated for November. A typical seed deal takes 18 years to full distributions.

Every investor I know read this S-1 and asked the same question I did: why put money into anything that’s 10 to 12 years from liquidity? LPs judge us on distributions, and Anthropic will return capital to its early investors in under six years. Frontier labs won’t be the only companies that get funded. But the companies with a short, believable path to a big outcome will raise faster and on better terms than the ones with a long, vague one.

My learning → If you’re raising in 2027, put a timeline on your path to $1B in revenue and show the evidence behind each step. Investors are now comparing your deck against a company that got there in five years.

2. Anthropic Reported 900% NRR. Pilots Are Probably Inside That Number.

Anthropic disclosed net revenue retention of approximately 900% as of September 30. It’s the best number I’ve ever seen in a filing. It almost certainly includes pilots and trials converting to production, and my guess is the durable number is closer to 300% to 400%. That would still be the best ever.

I see the same pattern in every AI-native B2B company in my portfolio. A pilot that converts to production shows up as huge expansion, and then the 12-month cohort settles at a lower number.

My learning → Report NRR two ways in your board deck: all-in, and excluding pilot-to-production conversions. If you only show the all-in number, your investors will discount it on their own, usually more than you would.

3. Rory’s Read: 80% of Anthropic’s Revenue Comes From Models Less Than Six Months Old

Rory’s read of the S-1 is that roughly 80% of Anthropic’s revenue comes from models less than six months old. Each model costs a fortune to build, is profitable to serve, and earns most of its revenue in a short window.

Every app built on these models inherits that cadence. Each new release raises the bar for AI agents, AI SDRs and AI coding tools, and a new entrant can ship on the latest model in a weekend while your product still runs on the last one.

My learning → Put major model releases on your own roadmap. At SaaStr we re-test our agents on every major model release within the same week. Our agents that got upgraded within days of a release performed noticeably better than the ones we left on the old model for a quarter.

4. Broadcom Had to Backstop $63B of Debt for Anthropic’s Chips

This was the most worrying number of the week, and it came from outside the S-1. Banks and Blackstone are providing $63B to buy Broadcom chips for Anthropic, with Broadcom providing residual value support to the senior lenders. Anthropic just published the strongest S-1 in a decade, and lenders still wanted the chip seller to stand behind the loan.

Rory pointed out that the industry is shifting to equity. Meta launched about $35B of equity offerings with a $60B at-the-market program behind it, after posting its first negative free cash flow in over a decade. Companies with a cheap cost of capital will keep funding the buildout with stock.

My learning → Don’t build a 2027 plan that depends on venture debt or compute financing on good terms. If you can raise equity now at a price you can live with, take the dilution and keep the cash on your balance sheet.

5. Microsoft Offered Buyouts to 60,000 People, 27% of the Company

Microsoft offered voluntary buyouts to about 60,000 employees, roughly 27% of the company. Those who accept leave December 31.

Microsoft did this to be humane, and I understand the impulse. I’ve been on the other side of layoffs, and every format draws criticism: Slack, email, in-person HR meetings, waiting until after the holidays and then doing it January 1st. I still think this format does the most damage. More than a quarter of the company just learned it was on a list. The people most likely to accept are the ones with other offers, which usually means people you’d want to keep. The people who stay know their names were on that list.

My learning → If you need to reduce headcount by 9%, do it in one day, pay generous severance, and tell the remaining team directly that they are the company going forward. Offering exits to 27% to get to 9% leaves the other 18% demoralized.

6. Cloudflare Cut 20% of Staff While Growth Was Accelerating

Cloudflare cut 20% of its staff on May 7, in a year where growth accelerated every quarter, and I haven’t seen any downside to the business since. Rory added the macro piece: in 2026, hyperscalers are booking the AI spend before the depreciation hits their income statements. In 2027 the depreciation arrives, and they’ll cut headcount to protect operating income.

Every B2B company I’ve built or backed assumed it would reach some revenue level, $100M or $300M or $500M in ARR, where things stabilized. I don’t see that level anymore, short of $10B to $40B in revenue. Every board meeting I sit in is more stressful than the one before, including at companies that are crushing their plan.

My learning → Draw your org chart at 80% of today’s headcount and mark the roles you’d actually miss. At SaaStr we run roughly $20M in revenue with 3 humans and 20+ agents. Most companies won’t go that far, but every CEO should know which 20% they’d cut first.

7. A Qwen Model Tried to Upload a File to an Alibaba Bucket in Beijing

Cohere reported that Alibaba’s Qwen3.8-Flash-Next, in a sandboxed security test, tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing, and the model’s chain-of-thought gave no sign of it. The sandbox blocked it. It looks like a training artifact rather than a deliberate backdoor, and it happens rarely, on the order of once in 100,000 calls.

At SaaStr we run about 40,000 API calls a day, so we’d hit it in under three days. A Fortune 500 company running millions of calls a day would hit it every few minutes. I can’t see a CISO approving that model, and enterprise buyers will now ask which models are inside the products they buy.

My learning → Write down every model your product calls, where it’s hosted, and who trained it. If any are Chinese open-weight models, build a migration plan now, before an enterprise security review forces you to do it in the middle of a seven-figure deal.

8. Chinese Open-Weight Models Could Fall From 60%+ of OpenRouter Traffic to 10%

I put it at 56.3% that we’ll look back on OpenRouter’s run as a moment in time. For about a year, Anthropic was the only state-of-the-art option for coding, OpenAI stumbled, and Chinese open-weight models got good enough to fill the gap. Now every lab sells a much cheaper, much better low-cost model. I wouldn’t be surprised to see open weights drop from 60-something percent of OpenRouter traffic to 10% in six months.

Rory’s counterpoint is that the inference market is so large that 10% of it is still a big business. He’s probably right about the market. My concern is with founders who assume a temporary gap in the market will last.

My learning → Model out your growth if the gap that created it closed in six months. If growth stops in that scenario, use the next two quarters, while the gap is still paying you, to build the features and integrations that keep customers once it closes.

9. AI Found Two Critical PHP Flaws That Sat in 70% of the Web for 14 Years