Four or five key inference providers add up to roughly $5 bn – $6 bn in run‑rate revenue.
The combined run‑rate revenue of OpenAI and Anthropic is probably north of $100 bn to $110 bn right now.
By dollars, inference accounts for less than 5% of total AI revenue.
Build‑to‑prod is the biggest category for AI‑software money, and the real upside lies in moving code into production.
CoreWeave and Oracle are B‑tier players, hypersensitive to debt loads and market shocks.
If your CapEx budget is $20BN, it is not enough that demand goes up. It has to go up by a lot every year.
My final question to the team would be: ‘How are we going to get at least a $1BN position out of this?’
The key inference providers add up to roughly $5B to $6B in run‑rate revenue, under 5% of OpenAI and Anthropic combined.
I get more messages about this podcast every week than anything else I do. This is the only show in tech you have to listen to every week.
Modal raised $750 M at a $15.75 B valuation.
I noted that when you are in a market that big, predicting with any precision the mix between closed tokens and open tokens is irrelevant.
I said we are investors in both Anthropic and Fireworks, the closed tokens and the open tokens, and both are going to be huge.
I highlighted that $100 M pre‑seed rounds are now “normal.”
I disclosed that Bessemer raised $1.75 BN for its early‑stage fund and $4 BN for its growth‑stage fund.
I explained that when we invested in Abridge at $5 M ARR we paid a $700 M–$800 M valuation.
I said Anthropic’s $15 BN round was backed while the company was roughly $100 M ARR.
I revealed that we convinced 4 LPs who initially turned down Bessemer to add $24 M of commitments.
I stated that Lagora was at $20 M ARR when we led a $1.8 BN round.
I emphasized that what I care about is the NTM basis – the next‑12‑month upside and whether the multiple feels reasonable.
I noted that VCs don’t want revenue – it’s a dirty word until a company hits $100 M ARR+ and becomes a clear breakout winner.
I called for a new generation of “pickers” – true artisans – to rise up in venture selection.
In late‑October 2026 I posted a series of X updates (dated 2026‑10‑19 – 2026‑10‑20) with @spdholakia that unpack the current AI‑venture‑capital environment, describe a trillion‑dollar “super‑cycle”, argue that capital is a moat, and lay out why I am investing in Anthropic.
The AI/large‑model market is entering a trillion‑dollar “super‑cycle” that is unprecedented in frequency, with only a handful of prior “magic” moments in 31 years of software history.
Massive compute‑intensive AI infrastructure is capital‑intensive, creating a barrier to entry; capital therefore acts as a moat that fuels rapid scale, brand building, talent attraction and a virtuous flywheel.
Future venture success hinges on being either a “giant global platform” that can consolidate many deals or an “extreme specialist” with deep differentiated expertise; firms in the middle face a tough environment.
I decided to invest in Anthropic because its compute‑intensive model aligns with the capital‑intensive consolidation dynamics of the AI super‑cycle and AWS’s fresh capital injection signals strong partnership potential.
Valuation multiples in AI are judged by upside potential over the next 12 months rather than current price; Bessemer’s recent deals illustrate this, e.g., Anthropic at ~150 × ARR.
Both open‑source and closed AI models will coexist; precise forecasting of the split is irrelevant given the market size, so Bessemer invests in both (Anthropic – closed, Fireworks – open).
With abundant capital, companies are expected to stay private longer, with a major IPO unlock anticipated in 2028 after a private‑market run‑up.
- $100 M pre‑seed rounds are now “normal.”
- Bessemer raised $1.75 BN for its “early” fund.
- Bessemer raised $4 BN for its growth fund.
- Anthropic valued at ~150 × ARR.
- Legora’s round: $1.8 BN at $20 M ARR.
We spend probably roughly 30% of our engineering salary on coding tools.
It has actually gone down the last two months, while the number of tokens that we have used has gone up.
First is visibility, second is basic cost controls, and third is we control the harness.
In that harness, we have done a bunch of work to control costs, things like prompt caching and routing.
I would still say it is below accuracy, but probably higher than latency.
For large enterprises, security is higher, but still probably below accuracy and cost.
Response rate: 94%
Response rate on group email: 18%
In mid‑October 2026 I dived deep into the emerging agent ecosystem, highlighting how both horizontal and vertical agents will coexist, why memory layers are poised to become ubiquitous, and the importance of neutral harnesses. I also called out the brutal reality of venture pricing and shared concrete numbers from LangChain that illustrate the scale of the market.
LangChain’s valuation today is $1.2 BN.
The company serves 35 % of the Fortune 500.
LangChain spends roughly 30 % of its engineering salary on coding tools.
Investor @milesgrimshaw joined Benchmark in September 2023 and led the LangChain deal.
Instinct and Muse have already nailed personalization and learning with memory layers.
I called venture pricing “totally f******.”
On 2026‑10‑16 I moved a coordination conversation to direct messages with @Edvin1Andersson to discuss details privately. The reply also served as a public summary of the status of “Harry Stebbings’ thing,” noting that the first batch of the funding round has 21 of 25 items funded, all by Tier 1 investors.
First batch funded: 21 out of 25 items, all funded by Tier 1s.
I note that roughly 20 decacorns (companies valued at $10 B) were created per decade over the last five decades, amounting to about 100 such firms in total.
In the most recent year, 42 decacorns were created – more than double the historic decade‑average.
There have been only 130 public tech companies ever worth $10 B or more since the dawn of American capitalism.
Haiku 5.5’s price is about 0.1 of Sonnet’s price, offering roughly a ten‑fold cost advantage.
I believe pre‑seed startups should not be required to produce 5‑year financial plans.
I argue that the world is changing so fast that long‑term forecasts are unreliable at the pre‑seed stage, and I call on investors to stop demanding 5‑year plans – this isn’t 1999.
Decision.delay_cost_optimization = true – I delay cost‑optimization until magical features prove product‑market fit.
Decision.avoid_own_model_claim = true – I avoid claiming a model is “their own” when it is fine‑tuned from a Chinese base.
Decision.exclude_chinese_models = true – I exclude Chinese‑origin models from production deployments.
Decision.leverage_cheaper_inference = true – I plan to leverage the expected drop in inference costs.
These decisions flow from the October X thread where I highlighted product‑market fit, model provenance, enterprise risk, and inference economics.
I reject the headless strategy = true. "Headless is brainless, because if you are giving up on your user interface and evolving your experience for a customer, I think that is a brainless bet."
I reject the magic‑button thesis = true. "I am a big bear on the magic button thesis of AI."
I avoid using Chinese models in production = true. "The biggest enterprises in the world are fearful of Chinese models."
I invest in inference cost reduction = true. "Inference will get so much cheaper over time. It will be wild."
I usually advise not to have a comp plan. I usually advise doing 100% OTE for at least the first two quarters.
You just do not know what you do not know. I have seen founders make a quota that gets blown out in the first six months and suddenly owe a seller multiple six figures.
We will pay 2X what a standard SDR is going to make, but you get to book 40, 50, 60 meetings, which is four to six times what a standard SDR has to book.
Over 200 after each show.
These points form the core of the "Moneyball" hiring framework I promote, emphasizing high‑pay, quota‑free onboarding, and targeting bankers as elite sales talent.
I overpay SDRs, paying them about 2× the standard OTE, which leads to reps earning 2X a standard SDR but booking 40, 50 or 60 meetings, four to six times the usual target.
Founders are advised to pay 100% OTE for at least the first two quarters instead of setting a comp plan.
Hiring one seller or SDR at a time leaves no way to benchmark performance, and comparing them against the founder is not reasonable.
I believe bankers make the best reps.
My recruiting “moneyball” target is top sellers at companies that grew revenue past nine figures despite a low NPS score.
These hiring and compensation principles came from a deep‑dive podcast with @max11free (SVP of Sales at @tryramp) that I promoted on X on 2026‑10‑10. The community’s enthusiastic replies reinforced the credibility of these ideas.
ElevenLabs is valued at $22 BN with a projected revenue runway toward $1 BN, and I say we go all in.
The durability of a company comes from a data loop where usage creates unique interaction data that improves the product, attracting more usage – a virtuous cycle that competitors cannot copy.
CROs jumping jobs frequently is wrong because it erodes the narrative and trust they have built with their teams.
In early October 2026 I posted a thread on X about CRO leadership ethics, tech‑week highlights, secondary‑market rankings and the rise of AI‑outcome engines. I argued that CROs should avoid jumping to direct competitors because it burns their reputation, while sector‑switches are more acceptable. I also shared concrete market numbers – OpenAI’s $70 BN run‑rate, Salesforce’s $2 B Listen Labs deal, Vercel’s $600 M ARR with agents now driving half of new business, and the SetterVC Q3 secondary‑market rankings.
CROs also hired other people to work for them, selling them on the company, the vision, and how it is going to be transformative for their family.
If that person were to leave, they have basically burnt their reputation with everyone they recruited as well.
I have never seen a CRO flip from one company to a direct competitor.
I think 95% of CROs would say, ‘This is how I found my last job.’
OpenAI Hits $70BN Run Rate.
Salesforce Buys Listen Labs for $2B.
Vercel has reached $600M in ARR, with agents now driving 50% of its new business, up from 3% at the start of the year.
SetterVC Q3 rankings: AnthropicAI #1, OpenAI #2, Etched #13, baseten #19, FireworksAI_HQ #16.
Defence sector jumps: HadrianInc +9 spots to #11, shieldaitech #7 (+5), Saronic #9 (+5).
Prediction‑market drops: Kalshi –16 spots to #24, Polymarket –24 spots to #30.
In the October 2026 X thread I reflected on three personal and professional themes: the mentors I’m grateful for, the moment I got the keys to my new Belgravia apartment, and why I think calling seed checks “option bets” is a harmful framing.
I value guidance from “some of the greatest people in our industry,” which continues to shape how I mentor founders.
I received the keys to my Belgravia apartment on 2026‑10‑07 17:36 UTC after weighing options in Dubai and Spain.
I criticized framing seed checks as “option bets,” arguing it feels like a direct attack on what founders and early‑stage investors build every day.
Anthropic round size: $4 BN.
Projected firm‑wide gain from Anthropic: $50 BN+.
Pay‑up rationale: investors may pay more because they see a bigger TAM than other investors.
Pay‑up can also be driven by the desire to win the deal.
What Menlo has done over the last 24 months is exceptional.
Yes, they did Anthropic’s round at $4BN that will make them $50BN+ as a firm.
Every seed investment is an option bet.
When you have 4X leverage, it does not matter if you are right.
Because LPs can buy index exposure to those same public tech giants without fees or carry, private VCs must generate outsized IRR performance, beating public markets by at least 1,000 basis points, to justify private capital allocation.
LPs cannot demand both smaller fund sizes and multi-year deployment paces in the middle of a generational platform shift.
Reflexivity creates self‑reinforcing loops where fast revenue growth triggers rapid valuation markups, attracting elite human and financial capital that further accelerates growth.
Premium entry prices are not always the result of irrational bidding.
Prediction: we will start to see the wave of customer support AI companies be acquired by incumbents.
Funding milestones: $42 M Series B led by @scalevp with participation from @NEA, @20vcfund, total funding now $65 M.
Namespacelabs Series B highlighted as a success in “agentic software development, CI/CD” infrastructure.
I continue to monitor these dynamics and adjust my fund's deployment strategy accordingly.
In early October 2026 I posted a series of X updates about Crusoe’s AI‑infrastructure strategy, highlighting vertical integration, contract structures, and broader economic impacts.
Crusoe is building an AI super‑major, vertically integrated across upstream, midstream and downstream.
The lead‑time for a key electrical distribution component dropped from 100 weeks to 28 weeks after internal manufacturing.
Capacity rentals are typically five‑year contracts with credit‑quality customers.
Short‑term services like Managed Inference and serverless fine‑tuning are higher‑margin than the longer‑term rentals.
Data‑center build‑outs drive a massive economic boom for blue‑collar labor in the United States.
Deploying data centers tends to lower local electricity prices by spurring new generation capacity.
In early October 2026 I shared a series of X posts covering AI market dynamics, funding norms, and my investment theses, adding concrete numbers on traffic shares, pricing, speed advantages, seed round sizing, and the landscape of Neo labs.
Jev is already 17% of the traffic on OpenRouter and 20% of the traffic through Vercel’s router.
Jev is a 70th of the price and 100 times faster.
Seed round should still be $2M to $3M.
There are still 100 Neo labs, so you have got to be in the 10 that win.
The best products are created by people solving a problem for themselves.
Based on these numbers I decided to allocate up to 30 % of the fund to Jev, maintain seed rounds at $2‑$3 M, and anticipate a wave of Neo‑lab acquisitions, while monitoring open‑source model economics.
For more context on my earlier observations see September 2026 posts.
On 1 October 2026 I posted on X about why I left Silicon Valley six years ago, calling the region a “monoculture” where every dinner party feels like a tech‑only gathering.
- I attribute my departure to a monoculture:
- I describe a stack‑ranking hierarchy that I see as pure toxicity that nerfs creativity:
I end the post with a self‑deprecating note: “That’s why I never moved. 😉”
These observations raise open questions about how pervasive this ranking mindset is across tech hubs and what alternative social structures might mitigate the monoculture effect.
- Founder of a company with a $100 B market cap
- Founder of a company with a $50 B market cap
- Managing partner of a tier‑1 venture firm
- Hired CEO of a $50 B public company
- Founder of a $10 B company
- …etc.
- “Because it's a monoculture”
- “Show up to a dinner party and everyone is in tech”