With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll
Come meet Rory live at SaaStr AI Annual. Rory O’Driscoll will be doing a live AMA on stage at SaaStr AI Annual on May 12 in the SF Bay Area. Bring your hardest questions on public B2B markets, AI capex, the SaaS reacceleration, and where the puck is going for venture in 2026. See you there.
This was the Super Bowl of earnings.
Five of the seven largest market cap companies on the planet reported in the same week. $540 billion in combined quarterly revenue. $700 billion in 2026 AI capex. And the punchline of the whole thing, which Rory borrowed from Evan Armstrong’s Substack: this is the most aggressive quarter in American capitalism.
- Google ran away with it. Cloud backlog nearly doubled to $462 billion, which is now larger than Alphabet’s entire 2025 revenue.
- Microsoft caught a downgrade despite $37B in AI ARR because excluding the AI initiative, the rest of the business is flat.
- Amazon got a thumbs up on AWS reacceleration.
- Meta got crushed for $145B in capex with no clear AI revenue line attached.
- And Apple quietly punched out a great quarter with no AI story at all and went home. The top of the distribution is pulling away from everyone else. That’s the macro frame for everything else this week.
Then Palantir reported on a separate planet entirely.
RPO up 134% to $4.45 billion. Rule of 40 at 145%, a number only matched in modern history by Nvidia, Micron, and SK Hynix.
Karp essentially saying every stakeholder in every commercial buying meeting now shows up, which is a level of compression he has never seen in his career. The reason is simple and brutal: corporate America has finally agreed that AI is the way to transform the company, and the only software vendor that can credibly take a $100 million bet to redo your go-to-market or business intelligence stack is Palantir. Everyone else moves in $200K chunks. That’s not a bet, that’s a feature purchase.
Underneath the mega cap drama, the real signal of the week was the SaaS reacceleration. Atlassian +29%. Twilio +20%. Five9 +23%. Palantir on its own continent. The B2B apocalypse narrative isn’t over, but it’s no longer the only story. The framework for who survives is now binary: monetize your existing base with AI AND attract net new customers. One prong without the other is a slow ice cube. Atlassian got prong one. Twilio got both. HubSpot just announced agents will be on par with humans in their next release, which is the right vision and arrives a little late but not too late.
Anthropic raised $50B at $900B in 48 hours by sending out an email. There is no IPO on planet Earth that beats that. Sierra raised $950M at $15.8B on $150M in ARR, a 105x multiple, which Rory framed correctly as the next-generation software counter-narrative to “LLMs eat everything.” The Musk vs Altman trial entered week one with a distillation admission and a $30B Brockman stake disclosure that nobody put any personal capital behind.
And Brian Armstrong said the quiet part out loud at Coinbase: build or go.
Anyone who can’t ship and manage at the same time is out. Anyone on LinkedIn talking about “my team” is out. Lead from the front with AI or step aside. Every founder secretly wants this world. He just made it the policy.
Top Takeaways
1. The Mag 7 Super Bowl: $540B in Revenue, $700B in Capex, and the Top Pulling Away
Forget the individual results for a second and zoom out a million miles. Five of the seven largest companies on the planet are accelerating at scale, with 20% topline growth and 30-40% growth in some subsegments. And they’re letting capex grow 50-60% on top of that, such that capex is now eating most of their free cash flow.
This is leaning in like nothing we’ve ever seen. Normally it’s the upstarts being aggressive and the incumbents defending turf. Here, six of the seven largest market cap companies on the planet (counting Nvidia) just said hell no, we’re not going to get pushed around, we’re going to make the bet too. The top of the distribution is pulling away.
The somewhat dirty subtext: a big chunk of what these companies are boasting about is that they sold a lot of compute to two privately held LLM companies, and that they bought some of those tokens and resold them through their distribution. Both statements are true. Both made the revenue line go up. But zoomed out: the five largest market cap companies on Earth are effectively working as distribution and capex providers for two privately held companies that ultimately own the IP. That’s an interesting structural setup.
The risk anyone running these spreadsheets hits is the same: if the AI bet is wrong, all these valuations are wrong. Microsoft’s existing business excluding AI initiatives is flat to slightly down. The growth is coming entirely from these initiatives. Three years ago that wouldn’t have been true. Today it is.
2. Google Won the Quarter, but the Real Token Story Is Underwhelming
Of the five, Google ran away with it. Cloud backlog at $462 billion, growing 80% year on year, with everything clicking. Search didn’t die. Advertising didn’t die. SaaStr’s own SEO is up 60% year on year, the highest ever. The AI search disruption story that everyone wrote 18 months ago hasn’t happened economically.
But here’s the controversial take Rory raised. Google boasted that Gemini token production went from 10 billion per minute in Q4 to 16 billion per minute in Q1. That’s a 60% lift on the most aggressive quarter in American capitalism. Anthropic likely 10x’d in the same period. Tokens probably grew more than that.
Translation: the most aggressive quarter in American capitalism is an underperforming quarter relative to the privates. Google is the best of the public LLM providers, and they’re losing the coding battle, which is where every dollar is currently flowing. Coding is the tip of the spear, the canary in the coal mine, the motherload job for AI. Google is nowhere on coding compared to Anthropic and OpenAI.
The structural challenge for the hyperscalers: their growth is coming from selling compute to LLM companies and reselling LLM tokens through their distribution. The actual LLM IP layer accrues to two private companies. Even Google, which has its own model, is significantly underperforming the privates on the most attractive segment. That’s the bull case for Anthropic and the bear case underneath every hyperscaler print.
3. Microsoft’s $190B Bet: Strip Out AI and Revenue Is Flat
The single most sobering statistic from the entire earnings cycle. Strip out Azure AI growth and Copilot growth from Microsoft, and the rest of the business is flat to slightly down.
Microsoft’s AI ARR is $37B. Their 2026 capex is $190B. That math only works if the AI bet keeps compounding. If it doesn’t, Microsoft is another B2B company trading at three times revenues. Welcome to our world.
Three years ago this wouldn’t have been the case. Today, all the growth is coming from these initiatives, and the valuation is fully dependent on them continuing to scale. That’s not a comment on whether the bet is right or wrong. It’s a comment on the leverage. These companies have made the AI bet load-bearing for their entire valuation structure.
The thoughtful counter is this: when Wall Street gives you permission to spend, you spend. The cash on the balance sheet is trapped if you can’t deploy it. There are moments in time when the public market lets you spend without punishing the stock, and you should spend every dollar. Bezos had to fight Wall Street for 20 years to spend on AWS. Satya doesn’t have to fight anyone right now. He should grab it. The day will come when growth slows and the permission disappears, and at that point you’re literally at the mercy of getting another half cent of EPS.
The risk is exactly what Rory called out: when the permission is there but the ROI isn’t, you get groupthink and bad capital allocation. Right now the allocation is working. The more you allocate, the smarter you look. That’s also one of the definitions of a bull market.
4. Meta Got Crushed Because Wall Street Can’t Build a Spreadsheet for “Vibes”
Meta crushed earnings. $56B in revenue. EPS of $10.44 versus $6.67 expected. And the stock got destroyed because they raised 2026 capex from $125B to $145B. Same week, Google spent the same money and got plaudits.
The difference is straightforward. Google has revenue coming in that’s clear and attributable. Meta doesn’t. Meta’s pitch is two-pronged: one, we’re optimizing ad performance using these models, and yes the math says they’re getting a 10-15% lift. But $145B in capex for a 10% lift on a $200B business is a brutal ROI. Two, we want to be present when people stop talking to other humans and start talking to chatbots, so we’re going to build next-generation experiences.
Wall Street can model the first one. They cannot model the second one. The second one is essentially a $150B bet on a future that isn’t articulated, and the analyst spreadsheets running GPU depreciation models can’t tie the math. So they put a higher discount on it.
The deeper truth is Mark isn’t running those spreadsheets and doesn’t care about them. He didn’t make $200 billion building a spreadsheet at Harvard, and he isn’t going to start now. He’ll spend to be relevant, and he might be right like Instagram and WhatsApp, or wrong like Meta itself. Either way, he doesn’t need permission. The thing is, when growth slows, that permission disappears for everyone else.
5. Palantir’s Rule of 145 and the Real Reason Big Companies Buy Big Software
This is the most important earnings story of the week if you’re trying to understand how AI buying actually works inside the Fortune 500. RPO up 134% to $4.45 billion. Rule of 40 at 145%, only ever matched by AI infra companies (Nvidia, Micron, SK Hynix). And Karp essentially predicted another year of doubling on the call.
Why does Palantir win? Rory’s framing was the cleanest. If you’re running application software, you want to be one of the top two initiatives for the most senior person you’re selling to. That’s how you make money. Right now in corporate America, the top two initiatives for every CEO are: launch the new product, and do something in AI. Every board is telling every CEO to get on top of this.
Then look at the buying menu. You can sign up Anthropic or OpenAI APIs and put Claude on every employee’s desktop. That’s individual productivity. Doesn’t move the needle. You can buy a point solution like Sierra for customer support at $2M. Or a smaller AI app vendor at $200K. None of those are enterprise transformation. They’re features.
Or you can hire the one company that has been moving in $20-100 million chunks for the US government for 20 years and credibly say: yeah, you want to redo your entire GTM stack with AI? Redo your business intelligence with AI? We can do that. We did it for the government. We did it for JP Morgan. CEO walks back to the board, marks the AI initiative as done, moves on to the next priority.
Big companies have to spend big money to do big things. There are very few software companies positioned for that motion. It’s why IBM existed for 30 years longer than it should have. It’s why EDS made a fortune as an outsourcer. Palantir is the only company on the planet right now that can credibly say to a Fortune 500 CEO: we will deliver enterprise-wide AI transformation in less than a year.
The Karp commentary on the buying motion was the most striking part of the call. He said for his entire career on the commercial side, he gets brought in by one stakeholder, sells to another, sells to a third, and it takes a couple of years. In the last year, every stakeholder shows up to the meeting. Everyone is there. The CEO and CFO are now telling everyone: come to the table now, we have to do this, we’re not evaluating for two years.
This is the COVID buying cycle of AI. Compressed, mandated from the top, with no naysayer voice in the room. Which means there’s going to be some serious capital misallocation downstream. But for the company that can actually absorb $20-100M of corporate intent and turn it into a deployment, this is the best moment in the company’s history.
At a $349B market cap, Palantir is priced to more than perfection. Two years of doubling and it looks cheap. Karp guided to roughly that. If the boom has legs, they’re the most likely to grow into the multiple. If it doesn’t, this one corrects hard. But the underlying enterprise dynamic isn’t going anywhere. No one has this expertise in-house. Nobody. It’s the worst gap between in-house and external expertise of our lifetimes, and it’s going to benefit Palantir for years.
6. The SaaS Apocalypse Counter-Narrative: Atlassian, Twilio, and the Two-Pronged Test
The reacceleration data this week was real. Atlassian +29%. Twilio +20%. Plus Five9 re-accelerated. Three older B2B companies, all of which were in some level of doghouse status six months ago, all of which printed numbers that a year ago would have been unimaginable.
But the right framework here is what Jason laid out: a two-pronged AI beneficiary test. Prong one: can you monetize your existing base with AI? Prong two: can you attract net new customers driven by AI?
Atlassian got prong one. Their Rovo AI product sold the f out of it, daily active users on AI features jumped, and AI revenues took a real step. But net new customer count is still slowing. That’s a one-pronged win, which means Atlassian is reaccelerating its base monetization but the long-run fate is still partially deferred.
Twilio got both prongs. Net new customer count may have grown 40% in the last year because every AI startup uses Twilio. ElevenLabs uses Twilio. Sierra runs on Twilio. The infrastructure is good enough that nobody bothered to rebuild it, and as agentic products scale, they default to it. Twilio is the cleaner story.
The third group worth watching is HubSpot. They announced this week that agents will be on par with humans in their next release, with their platform fully open to agents. That’s the right strategic bet. It’s a little late. It’s not too late. If HubSpot can become the hub for agents in all of its categories for SMBs and GTM, they should reaccelerate dramatically. If it works, the playbook is clear for every classic B2B company. If it doesn’t work at HubSpot, given how broad their customer base is, you can write off most of the rest of the category.
The clearest framing is this: reacceleration is going to be the exception, not the rule. The classic B2B companies that don’t get to both prongs become slow ice cubes, growing in the 10s instead of evaporating, but with no path to a re-rating. The ones that hit both prongs get to a 30%+ reaccelerating cash-flow-positive profile, which historically gets a 6x revenue multiple. That’s the actual destination, not LLM-style 100x dreams. And ironically, what Atlassian and Twilio just did becomes the template that exposes everyone who can’t do it.
7. Anthropic at $900B: The 48-Hour Raise That Killed the IPO Argument
Anthropic raised $50B at a $900B valuation in 48 hours. No drama. Email out, terms accepted, money wired. There is no IPO on the planet that competes with that.
This was the correction Rory made openly: two weeks ago he thought Anthropic should skip the round and go straight to IPO. Wrong call. In the world we live in now, you can raise $50B in two days with no IPO blocks, no rights triggers, no statutory liabilities, no disclosure requirements, no roadshow. Why on Earth would you go public if you can do that? You wouldn’t.
The math underneath the raise is what matters. Every dollar of revenue Anthropic does requires roughly $3-4 of capex from Anthropic or its hyperscaler partners to service. And because they’re going 10x year on year, they have to forecast that capex one year out. So at $10B in run rate revenue, you’re committing $30B+ in capex against a $1B revenue base today. Across the whole stack, this is the riskiest game of financial guesswork ever played. Anyone betting 5-10x their current revenue on capex to meet next year’s demand has never existed before at this scale.
The only thing you can do is derisk the bet, and the only way to derisk is to raise capital. There is no such thing as too much cash on the balance sheet right now. Dario is entirely right.
Also worth watching: this raise probably pushes the IPO further out. If OpenAI is now looking like a 2H 2027 IPO, and Anthropic is fully funded for the next 18 months, neither of them needs to deal with public market headaches in 2026. The only reason to go would be a particularly favorable market window, and even then, it’s optional rather than required.
For comparison: $250 million is now the entry price to take a meeting with the GPs of the leading firms as an LP. That tells you everything about where the demand sits.
8. Token Spend Per Engineer: The Number That Tells You How Big AI Actually Gets
Buried in the Anthropic discussion was the most useful framework of the episode for sizing the AI opportunity. The right question is: what’s the steady-state token spend as a percentage of salary dollars per engineer in a fully mature AI-first organization?
In coding, 20% token-to-salary ratio means Anthropic can grow into hundreds of billions in revenue. At 5% it gets a lot harder. Karpathy uses Claude for the final 20% of his coding work, which suggests the upper bound is real but the steady-state percentage is very much in motion.
But here’s where Jason punctured the assumption. He pulled the actual token spend numbers for SaaStr’s two production AI agents. QBee, the AI VP of Customer Success that manages 100+ sponsors and reaches out to 232 of them at midnight. 10K, the AI VP of Marketing that runs daily GTM standups, generates three better-than-human ideas every morning, and has started running its own campaigns. Combined token cost: $254 a month. For both. Together. Full-time autonomous.
For context, this replaces real human work that would have cost 15x that on salary alone. The 10K monthly token spend is $94. For an entity that beats every human marketer SaaStr has had on ideas. So if the world is converging to engineers spending 20% on tokens, but agentic workflows for marketing and customer success are running at sub-1% token-to-output ratios, the math for Anthropic looks different depending on which job you’re talking about.
There are two ways to read this. Bull case: tokens are dropping in price 10x every 18 months on hardware plus model optimization, so even if current efficiency is sub-1% in some categories, the absolute spend can grow as use cases multiply (continuous code reviews, meta-reviews, real-time agent fleets). Bear case: outside coding, AI may be more deflationary than anyone modeled, which means the dollar TAM for token providers in non-coding categories is smaller than the headcount-replacement story suggests.
Either way, this is the question to track. Token spend per FTE-equivalent, per category, per quarter. That’s the chart that explains whether Anthropic is going to $200B in revenue or $50B.
9. Sierra at $15.8B: The Counter-Narrative to “LLMs Eat Everything”
Sierra is raising $950M at $15.8B on $150M in ARR. That’s a 105x revenue multiple. Brett Taylor as CEO. OpenAI chairman. Every sophisticated LP in the world calling Harry asking how to get in.
The bull and bear cases are both worth taking seriously.
Bear case (Jason): the customer service software market is roughly $20-30B today. The customer service labor market is $400B. The implicit pitch at 100x revenue is that Sierra captures a meaningful slice of the $400B labor pool, not just the existing software pool. That requires real labor replacement plus expansion into sales and upsell. Possible, but at 100x, you’re underwriting both. And there are three or four other companies competing for the same outcome with the same LLM access. Your competition isn’t labor, it’s three other VC-backed startups with the same GPT or Claude API key.
Bull case (Rory): the existence of this round is itself the most important data point. Brett Taylor is the chairman of the largest LLM company on the planet. He clearly believes there’s a software value layer to be built on top of LLMs. So does Karp. So does every operator who has actually deployed agents inside a real enterprise. The “all software is dead because LLMs eat everything” thesis isn’t being voted on by analysts. It’s being voted on by dollars. And the dollars are saying that next-gen software companies built on top of LLMs are going to be massive.
The token intensity check matters here too. Sierra’s LLM cost as a percentage of revenue is likely sub-10%. So 90%+ of their value is the software layer plus domain knowledge plus deployment plus reliability, not the LLM itself. If that’s correct, the value capture is in the application layer, not the model layer. That’s the entire investment thesis for next-gen B2B + AI.
If you had a dollar to put to work between Sierra and Anthropic at current valuations, both Jason and Rory said Anthropic without hesitation. The Anthropic upside case is boundless. The Sierra case requires you to underwrite labor replacement AND fend off three credible competitors. Different bets, different risk profiles.
10. Apple’s Quiet Quarter and the Stealth Memory Inflation
Easy to miss in the noise. Apple beat across the board. Tim Cook punched out on a high. Stunning results, great quarter. No real AI story. No participation in the capex hysteria. Just continued buybacks and shareholder returns. Worth noting in the context of every other Mag 7 burning $100B+ on infrastructure: Apple is the one defendant in the room running the opposite playbook and printing money.
The more interesting subplot is memory chip inflation. A significant portion of the capex raises this quarter aren’t actually buying more physical capex, they’re buying the same amount at a higher price because memory costs have exploded. Apple just dropped the $599 Mac Mini and replaced it with a $799 SKU because of memory costs. That’s not a $200 actual cost difference, it’s pricing power being passed through. This is going to filter into iPhone pricing next year, and across the entire device industry. Stealth inflation that nobody is fully modeling.
11. Musk vs Altman, Week One: Soap Opera With Real Legal Stakes
Week one of the trial delivered TMZ-level entertainment. Musk admitted under oath that xAI distilled OpenAI models partly. Greg Brockman’s stake disclosed at $30B with zero personal capital invested and personal angel positions in Cerebras (now a $10B OpenAI customer) and Helion. Personal diaries entered into evidence. Musk asked under oath to rank the LLM models, having to put OpenAI above Grok, which probably hurt deep in his soul.
Underneath the spectacle, the real legal questions are technical. One: statute of limitations. Did Musk wait too long to bring the case? If the judge finds he should have known the alleged harm earlier, the merits never get ruled on. Two: standing through the donor-advised fund. Most of Musk’s OpenAI funding flowed through his DAF. Once money goes into a DAF, it’s no longer his money. The DAF is a separate legal entity. So the DAF is the harmed party, not Musk personally. He may not have standing.
Probably worth watching for the entertainment value. Probably won’t change the outcome. The judge decides most of this regardless of the jury, because the jury is advisory in this case. Both sides will look bad on the stand because both sides are partially in the wrong. The real outcome will turn on technical legal issues that have nothing to do with the optics.
12. Brian Armstrong’s “Build or Go”: The End of the Manager-of-Managers
This is the most consequential operator story of the week. Brian Armstrong announced Coinbase no longer wants any employee who isn’t also an individual contributor. Managers who can’t ship are out. Managers of managers are out. If you can’t deliver a campaign and be the head of marketing, if you can’t manage sales agents and be a sales manager, you don’t work at Coinbase.
Every founder secretly wants this world. We’ve all watched the same playbook for 20 years: first 50 employees, everyone ships, everyone builds, founder knows everyone. Cross 500 and people start leaving before you’ve ever met them, as Aaron Levie famously discovered at Box. The capitulation to managers of managers happens around 200-500, and it has been the standard B2B operating model ever since.
Armstrong is the right person to break it. He has demonstrated clarity of thought before, most notably during the political-at-work moment in 2020 when he told everyone to leave politics out of the company and offered severance to anyone who disagreed. That call took heat at the time and turned out to be entirely correct. Coming from him, “build or go” is more credible than from a CEO using AI as a layoff excuse.
The deeper logic Jason laid out: the best executive in 2026 is one who can interact with the agent directly. A CMO today should be able to spin up campaigns themselves through an agent, not through a marketing director through a campaign manager. SaaStr’s AI VP of Marketing started running its own campaigns the last two weeks. Which means the CMO who insists on a team of three to do what the agent does autonomously is creating drag, not value.
The 10K monthly token spend that runs this. $94 a month for an autonomous AI marketing executive that beats any human SaaStr has had on idea generation. That’s the new operator math. If a manager can’t keep pace with that economics, the manager loses.
Maybe 5-10% of current managers can actually build alongside their agents. That’s the pool. The rest will rotate out, either because they self-select or because companies like Coinbase make the policy explicit. As Rory framed it: if the people who can do this succeed, capitalism is Darwinian and the ones who can’t will be forced out. The 10% becomes 20%. Then 40%. The standard moves.
This is the real AI labor reset. Not “AI replaces jobs.” More precise: AI eliminates the manager of managers tier in companies that take advantage of it, and elevates the operator who can ship and lead at the same time. That’s the entire memo.
Quotable Moments
Jason Lemkin
“Anyone on LinkedIn that talks about their “team”, fire them. My team this, my team that, they’re all so precious about their team. That means they did nothing. Lead from the effing front with AI.”
“If you’re a $20 billion plus public company software company, you’re dead. You don’t have permission from Wall Street to spend 10% to decrease your gross margins so you have the best agent in your category. You’re trapped in a death spiral.”
“$254 a month to run two highly valuable autonomous AI agents for marketing and customer success. Pretty crazy. Amelia literally thought that was per day.”
Harry Stebbings
“You’re right, $50 billion in 48 hours with no rights and no anything. There is no IPO on planet Earth that beats that.”
“Of the four. Alphabet, Amazon, Meta, and Microsoft. You can buy one and sell one. What do you buy? What do you sell?”
“I’ve had like seven LPs ping me being like, Harry, how do we get in? How do we get chairs? How do we get in?”
Rory O’Driscoll
“The most aggressive quarter in American capitalism. Five of the seven largest market cap companies on the planet saying hell no, we’re not going to get pushed around. We’re going to make the bet too.”
“Big companies have to spend big money to do big things. Palantir is like general catalyst for AI transformation. You need to move $100 million. You probably won’t get fired. Cool. Absolutely.”
“There is no such thing as too much cash on your balance sheet right now. This is the riskiest game of financial guesswork I’ve ever seen. The only thing you can do is derisk the bet, raise capital.”
This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll.
