Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/SOURCERY NEWSLETTER/BREAKING: Altimeter’s ‘Four Seas…
NEWS
// NEWSLETTER ISSUE
SOURCERY NEWSLETTER

BREAKING: Altimeter’s ‘Four Season of AI’

DATE November 21, 2026SOURCE SOURCERY NEWSLETTERPARTICIPANTS MOLLY O'SHEA
// KEY TAKEAWAYS5 ITEMS
  1. 01AI is a historic capital supercycle that is still early
  2. 02The shift from chatbots to agents is changing what AI is used for
  3. 03A power law in usage creates both a gap and an opportunity
  4. 04Compute ownership and scarcity are the new competitive moat
  5. 05The growth benchmarks for AI companies have been reset
// SUMMARY

1. Key Themes

AI is a historic capital supercycle that is still early

"It's about $860 billion this year, over tripled in the last 2 years. And to put that in perspective, it's now rivaling the US defense spend."

AI CapEx now exceeds the share of GDP of the railroad, oil and gas, telco, and shale booms. The value chain is also front-loaded toward semiconductors, which follows the mobile-era pattern of value flowing to chips, then infrastructure, then apps.

"While the economy has grown 6X in the last 2, 3 years, the apps and model side of it has grown about 100X in the same time."

In cloud, application revenue (~$600B) dwarfs infrastructure (~$300B) and semis (~$80B). In AI, models and apps make only ~$140-150B on a much larger semiconductor base, implying the application layer has room to catch up.

The shift from chatbots to agents is changing what AI is used for

"Sometime end of last year, the start of this year, we entered what I call the agentic AI era. These are long-running agents, swarms of agents, and the primary use case for those is very different. It's coding, agentic automation, and by the way, this is not just for technical users. This is for sales, marketing, support teams, investment teams."

About 80% of chatbot use is knowledge retrieval, while coding and automation make up about 80% of agentic use. Apoorv expects a third phase, with proactive agents, longer-running swarms, and large models performing surgery on smaller ones.

"I don't think this is the last phase shift we'll see. I'm sure there'll be a phase 3 of AI that people use."

A power law in usage creates both a gap and an opportunity

"Over the last year, year and a half, top 10% of their customers have grown their usage 17x, while their median customer has grown their usage 2x. And so the gap between a frontier firm and a median firm using AI is widening."

At Cursor, the top 10% of users account for two-thirds of all tokens. Apoorv frames the gap as upside rather than a risk:

"There's a power law in usage, but that is what creates the opportunity for a catch-up for the rest of the population as AI adoption grows, and I would say that's the biggest gap in AI right now, which is not everybody is using it like the power users are."

Compute ownership and scarcity are the new competitive moat

SpaceX rents compute at spot prices far above neocloud lease rates, and still recovers its CapEx in under a year. Neoclouds sell 1-3 year leases at roughly $10-20 a watt, while SpaceX rents at $25-50 a watt. Anthropic can pay those rates because it monetizes compute at $50-100 a watt.

"It's like renting your home on a long yearly lease versus renting your home on Airbnb. There's obvious questions with that, whether Anthropic will renew and so on. But I guess we'll find out."

On supply constraints, Altimeter expects 10-15 GW less capacity than forecast next year:

"What happens to the neoclouds? What happens to new startups that were, let's say, not the top 5 in the line, but let's say 17th in the line? I think for them, this is existential. For them, this is a much bigger question than it is for the hyperscalers."

"I think the answer is chips are going to win, data centers are going to win, clouds are going to win, models are going to win."

The growth benchmarks for AI companies have been reset

"What was true in software, which is that you could triple, triple and double, double, is no longer best in class... It's 10X at single billion, 3X at double-digit billion, and maybe doubling at triple-digit billion is probably best in class."

Cursor is rumored to have gone from about $1B to $10B this year, Baseten has grown 10X in each of the last 2 years, OpenAI is expected to go from ~$20B to $70-80B, and Google Cloud is growing over 100% at triple-digit billions.

2. Contrarian Perspectives

Open source is "evergreen" and monetizes through serving infrastructure, not model licensing

Where many see open source as a free commodity threatening the labs, Apoorv argues the value accrues to the serving layer. Open source runs about six months behind the closed frontier and has held that gap steady. His mental model is that 20-30% of AI spend will go to open source.

"Open source is not free. As open source, you still have to go build the cluster that'll serve your model. And that's, we believe, the primary place a lot of gross margin will accrue."

Evidence: Groq CEO Jonathan Ross says about 35 companies run 99% of inference, and the top 5 are the labs. The next 30 are AI-native startups, the fastest adopters of open source tokens. OpenAI's partnership with Baseten lets customers spend commitments on open source models. Apoorv expects Baseten, Fireworks, and Modal to share the market like AWS, GCP, and Azure.

Don't forget Google, and Meta is not in the frontier

While the narrative centers on Anthropic and OpenAI, Apoorv argues Google is underappreciated, with a search business that is slowing but a cloud business growing 8X faster.

"On one side, they've got a search business that is slowing down... But you've got this cloud business that is half the size of search, but growing 8X faster."

"Every time I think of Google, I say, 'Don't forget Google.' ... It's probably a change of guard and change of seasons inside of Google."

On Meta, when asked about its Muse model, Apoorv kept the frontier to Anthropic and OpenAI.

A $1.5T Anthropic valuation is "not heroic"

Even as Anthropic fell 8.5% on its second day amid a tech selloff, Apoorv defends the price. He put the valuation at 11-14x forward GAAP revenue, on his estimate of $140-150B revenue next year.

"It's not a heroic multiple. It's completely reasonable based on current market comps."

"This is the private Mag 7. This is likely to be the leading companies for the next decade and more."

Cyber will be won by frontier labs, not standalone startups

"The edge that you get in using frontier models is so high that I think it'll be the frontier companies. Today that's OpenAI and Anthropic that are likely to serve that demand."

Apoorv notes that no standalone cyber company has broken out yet, and that companies are moving from annual penetration tests to continuous testing as they ship software.

3. Companies Identified

Anthropic

  • Description: Frontier AI lab, maker of Claude; recently IPO'd
  • Why mentioned: Leader at the frontier and the most capital-efficient scale-up Apoorv has seen
  • Quotes: "This is a very rare sauce, like I've never seen this before, a business that's growing this fast with no customer concentration and a long runway to grow." Also: "They burned less than $10 billion to get to that number, and I mean, that has got to be the... like the books of history." (10X annual growth for 3 years, ~$17B Q3 revenue, ~12,000 customers over $100K, compute cost per $1 of revenue down from 88 cents to 15 cents.)

OpenAI

  • Description: Frontier AI lab behind ChatGPT, Codex
  • Why mentioned: Altimeter's largest investment; rebuilt its sales organization and is becoming a platform
  • Quotes: "I think about 2 jobs to be done, building and selling. On the building side, they've had an incredible set of models that have come out... that have defined the frontier." Also: "I think it's very smart of OpenAI to rise above the level of being a lab to becoming a platform."

SpaceX

  • Description: Public company (Nasdaq: SPCX) spanning launch, Starlink, and AI compute
  • Why mentioned: Now an AI infrastructure business, renting compute at premium spot prices
  • Quotes: "We thought SpaceX was a space business, but turns out it's an AI business." (~50% of revenue from AI; Google pays $920M per month for ~110,000 Nvidia GPUs.)

Google / Alphabet

  • Description: Search giant with a fast-growing Cloud and TPU business
  • Why mentioned: One of the four seasons; Cloud growing 103%
  • Quotes: "...you've got this cloud business that is half the size of search, but growing 8X faster."

Baseten

  • Description: Inference platform and Altimeter portfolio company
  • Why mentioned: Positioned to capture open source inference spend; adding routing and sandboxes, moving toward a hyperscaler model
  • Quotes: "You now have the big labs, OpenAI, making the choice that, 'Hey, you can take your spend on OpenAI and spend it on open source models.'"

Fireworks and Modal

  • Description: Inference platforms
  • Why mentioned: Expected to share the inference market with Baseten, as AWS, GCP, and Azure share cloud
  • Quotes: See the Baseten entry; the article names all three together as positioned to capture value as the AI economy expands.

Cursor

  • Description: AI coding product
  • Why mentioned: Rumored to go from ~$1B to $10B this year; top 10% of users drive two-thirds of tokens
  • Quotes: "Cursor is rumored to have gone from about $1B to $10B this year."

Cognition

  • Description: AI coding company
  • Why mentioned: On a growth curve similar to Cursor
  • Quotes: "Cognition is on a similar curve."

XBOW

  • Description: AI cybersecurity company, Altimeter portfolio company
  • Why mentioned: Its team found a Google Chrome bug allowing computer takeover with a single link click
  • Quotes: "Offensive actors now have this new toolkit, which is incredibly potent, that can go check every single door and window and make sure if it's unlocked, find the most critical vulnerability, and exploit it."

CoreWeave and Nebius

  • Description: Neoclouds
  • Why mentioned: Sell 1-3 year leases at ~$10-20 a watt; most exposed to the capacity shortfall (CoreWeave fell double digits in the selloff)
  • Quotes: "A typical neocloud would let you buy a long-term lease for 1-3 years. SpaceX is renting this at spot prices."

Altimeter Capital

  • Description: Investment firm
  • Why mentioned: Co-led Anthropic's Series H; OpenAI is its largest investment; planned $5B in the Anthropic IPO
  • Quotes: "Altimeter co-led Anthropic's Series H, & OpenAI is the firm's largest investment."

Other portfolio and mentioned companies: Glean and Parloa (Altimeter investments led by Apoorv); Nvidia, Amazon Trainium, Google TPUs, and Broadcom (Anthropic's compute); Oracle (first outside cloud to offer TPUs); Micron, Sandisk (fell in the selloff); Palantir (Apoorv's former employer).

4. People Identified

Apoorv Agrawal

  • Description: Partner at Altimeter leading AI and software investing; former Palantir engineer; teaches a Stanford course on the economics of the AI stack
  • Why mentioned: Interview subject and creator of the "Four Seasons of AI" framework
  • Quotes: "I won't count out open source, which is kind of the evergreen season."

Molly O'Shea

  • Description: Host of Sourcery
  • Why mentioned: Interviewer and author
  • Quotes: "I sat down with Apoorv Agrawal, AI Partner at Altimeter, to break down Anthropic's financials, the economics of the AI supercycle."

Brad Gerstner

  • Description: Altimeter Capital leader
  • Why mentioned: Said Altimeter planned to invest $5B in the Anthropic IPO, five times his prior IPO high
  • Quotes: "Brad Gerstner said Altimeter Capital planned to invest $5 billion in the Anthropic IPO, which is five times the most he has put into any prior IPO."

Elon Musk

  • Description: Founder of SpaceX, Tesla, xAI, and others
  • Why mentioned: Made the compute bet early; praised for mastering atoms and bits; coined "zero day, every day"
  • Quotes: "Elon's probably the best entrepreneur we know today in orchestrating atoms."

Sam Altman

  • Description: OpenAI CEO
  • Why mentioned: Named alongside Musk as a founder who made the compute bet early and got it right

Jonathan Ross

  • Description: Groq CEO
  • Why mentioned: Source of the claim that ~35 companies run 99% of inference
  • Quotes: "Groq CEO Jonathan Ross has said about 35 companies run 99% of all inference."

Bob McGrew

  • Description: Former PayPal and early Palantir engineer, helped develop the forward-deployed engineer model, former OpenAI research lead, now a founder backed by Altimeter
  • Why mentioned: Example of talent as a leading indicator
  • Quotes: "The guy's got a track record of just picking the most ambitious bets. I think it's talent like that that is a leading indicator."

Michael Truell

  • Description: Cursor founder
  • Why mentioned: Cited as a founder example for the "founders as leading indicator" thesis

Oege de Moor

  • Description: XBOW founder
  • Why mentioned: His team found the Chrome vulnerability

Alex Karp

  • Description: Palantir CEO
  • Why mentioned: Launched the "AI Sovereignty" movement referenced in connection with open source adoption

Shaun Maguire

  • Description: Investor quoted in a linked clip
  • Why mentioned: Argues xAI is underrated
  • Quotes: "I think xAI is severely underrated."

5. Operating Insights

Reset your growth benchmarks for AI-era scale

The software "triple, triple, double, double" playbook no longer defines best in class. Operators should benchmark against 10X at single-digit billions, 3X at double-digit billions, and ~2X at triple-digit billions.

"It's 10X at single billion, 3X at double-digit billion, and maybe doubling at triple-digit billion is probably best in class."

Own your intelligence by fine-tuning smaller models on customer data

For AI-native companies, using customer interactions to fine-tune a smaller model lowers inference cost and builds a data flywheel.

"You own that intelligence as opposed to renting that intelligence."

Secure compute early, and diversify across chip suppliers

With a potential 10-15 GW shortfall, startups far back in the queue face existential risk. Anthropic's multi-chip approach (Google TPUs, Amazon Trainium, Nvidia GPUs) is the model for reducing dependence on one supplier, while the SpaceX example shows premium spot capacity is available if the unit economics support it.

"I think for them, this is existential."

6. Overlooked Insights

Frontier customers see AI as a tiny share of their budgets

Fortune 500 customers spending over $100K on Anthropic say it is less than 0.1% of their total expenses, a sign of how little budget share AI has captured relative to its capabilities.

"Fortune 500 customers that spend more than $100K on Anthropic say it is less than 0.1% of their total expenses."

SpaceX's growth targets depend on customer renewals at premium prices

SpaceX's targets for next year assume about 13 GW of compute monetized at or above current rates. That makes the business sensitive to whether customers like Anthropic renew at spot-style prices, and to whether supply constraints persist to support pricing.

"SpaceX's targets for next year assume about 13 GW of compute monetized at or above current rates, so renewals by customers like Anthropic matter."