With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll
Apple sued OpenAI for trade secret theft this week, and the details are ugly: a six-year Apple employee allegedly walked physical prototypes out the door for show-and-tell, encouraged by a 24-year Apple veteran now running OpenAI’s hardware group. The individuals involved are, in Jason’s words, toast. But the more interesting read is what the lawsuit signals about OpenAI’s hardware ambitions, which look increasingly like a distraction from the one thing actually printing money right now: coding.
That tension ran through the whole episode. Meta re-entered the frontier race by launching Spark 1.1 and, for the first time ever, charging developers for its own models. Zuck broke a three-year silence on X to do it. The battle is aggressively priced and aimed squarely at the cheap-token tier, because every company with a half-awake CIO is about to hand its developers a budget model to stop the bleeding.
And the bleeding is real. ClickHouse’s AI spend is up 60x since February. Jason burned an entire month of Claude Design credits redesigning a single page. The best developers are now coding 24 hours a day across 10 to 20 agents, and none of them can consume enough tokens. Which raises the question Rory kept circling: there are only 1.8 million developers in the US and roughly $250 billion in total developer wages. If most of Anthropic and OpenAI’s enterprise revenue is coding, they may already be at 20% of the entire software labor market. Nobody on the pod had a conclusion, but the possibility is real: the fastest-growing companies in history could hit the ceiling of the till faster than any company ever has.
Top Takeaways
1. Apple vs. OpenAI Won’t Slow the Roadmap, But It Reprices AI Talent
The individuals are done. There’s fact-based evidence in the complaint, the junior employee who took the material has likely already lost his job, and the first thing he’ll learn, as Rory put it, is that there’s no gratitude in litigation. The moment it’s convenient, OpenAI burns him.
Jason read is that the company-level story is leverage. Apple is furious that 400 of its people have gone to OpenAI, and now it has the junior guy, probably the VP of engineering, and a discovery process it can use to march up the chain. None of that likely moves OpenAI’s ship dates. It’s a message, and the audience is every engineer thinking about making the same jump: Apple is going to make poaching expensive, publicly, at least once.
2. California’s Employment Law Made the Theft Pointless
The theft wasn’t just wrong, it was unnecessary. California doesn’t enforce non-competes, and the doctrine of inevitable disclosure means the knowledge in your head travels with you legally. You can hire the domain expert and use what they know, which makes taking on criminal risk to steal it pure downside.
Anthropic is the proof case. When those founding employees left, they didn’t need a line of code or a piece of paper, because everything in their brains was already theirs. Rory’s larger point is that California’s stance is a rare policy triumph: it’s terrible for the individual employer trying to hold people, and it’s the single biggest reason the Bay Area keeps out-innovating everywhere with tighter employment law. The moat in AI is people, and the law is built to let people move.
3. OpenAI’s Hardware Bet Looks Like It’s on the Cutting Board
This one comes down to focus. Hardware made sense when OpenAI had an unassailable consumer lead and everything was working. It makes far less sense now that the value has concentrated so violently in one place: enterprise coding. Sora’s been trimmed, the TBPN deal looks questionable, and the $6 billion for Jony Ive’s team has so far produced a lawsuit instead of a product.
Jason’s read is that this could be the excuse that finally kills it, the way a bad quarter gives a CEO cover to cut the project everyone privately knew was a distraction. Rory pulled in Ben Thompson’s framing that consumer success may itself have pulled OpenAI away from where the money is. As a consumer company, hardware and media are logical concentric circles. As a company that now knows the real economics are in coding, they’re a death march. Watch for OpenAI to narrow.
4. Meta’s Spark Launch Is a Price War for the Second Tier
The benchmarks matter less than the business-model shift: Meta moved off open weights and started charging on an API, which means the last major holdout just adopted the same model as every other frontier lab. Alexander Wang called the pricing very aggressive versus OpenAI and Anthropic. This is a low-priced entry aimed at the cheap-token bucket, not the frontier.
Jason found it competitive with Claude on some consumer answers but flagged the real test is coding, where Meta’s benchmarks have overpromised before. The open question is whether it’s ever ROI-positive to be the fourth player selling at that price. Being aggressive on price is easy when you have Meta’s balance sheet. Being right about it over five to ten years is a different bet.
5. Every Company Is About to Run a Two-Tier Model Strategy
Token-maxing is what forces this. Everyone hits their limits, every organization moves from “burn all the tokens you want” to budgets, and the moment budgets arrive, you need a cheap model for the simple work and an expensive one for the hard work. It’s an operating constraint, and it creates a permanent, high-volume slot at the bottom.
Haiku is Jason’s example of how quickly that slot can be filled: excellent for simple work at a tenth of a cent, and improvable overnight by dialing up how much Opus reasoning flows into it. The frontier fight gets the tweets. The volume, and a lot of unresolved questions about where the margin actually is, sits one tier down.
6. Cost Per Completed Task Kills Cost Per Token as a Buying Metric
Every CIO is about to start buying this way, and the Databricks paper Rory broke down is why. A model with cheap base tokens can carry expensive reasoning tokens and wildly unpredictable consumption, so it blows past a nominally pricier model on the only number that matters, which is what it costs to actually finish the job. Price-per-token marketing is now borderline misleading.
Two consequences follow. Different models win different tasks, so you get a Pareto curve rather than a single winner, and the harness around the model, the infrastructure and orchestration, becomes a real value center rather than plumbing. Rory made the caveat himself: Databricks sells model management, and here’s a Databricks paper arguing model management matters. They also brought the data, and it lines up with what operators are seeing, which is where Jason pointed to ClickHouse’s 60x spend increase since February.
7. The Richest Token Workflows Haven’t Even Started
Consumption is nowhere near saturation, and the reason is that outputs keep getting richer faster than tokens get cheaper. The illustrative version: instead of redesigning one page, have Claude Design redo your entire 100-page site every night and pick the two or three better versions in the morning. That’s not a rounding error more tokens, it’s orders of magnitude, and it’s a better outcome. Every strong developer and designer could plausibly consume 100x what they do today.
Underneath the enthusiasm is a governance problem Rory named cleanly. There has never been a product that does a worker’s job for them, on demand, at no personal cost, while the company silently absorbs the expense. Often that’s wildly accretive, twenty dollars of tokens saving five hundred dollars of labor. Without a governor, you eventually cross the line where it’s six hundred dollars of tokens to save five hundred. Every finance team is about to go looking for that governor.
8. For Incumbents, the AI Threat Is Slow Funnel Erosion
This is Jason’s ankle-biter thesis. Claude Design won’t kill Figma and won’t dent a quarter, because what it takes are the single-seat deals at the very bottom. The danger is slower and structural: those ankle-biters move up the leg. If the next generation starts agentically and never needs you, they never graduate into you, and the loss shows up years later as a funnel that quietly stopped refilling.
Salesforce carries the same exposure. The one number Jason watches on a public B2B company right now is net new logos. Above 15% a year, they’ll figure the rest out. The moment that growth cracks, price increases buy you a little time and then the death spiral starts, because you can’t raise your way out of a funnel that stopped converting new entrants.
9. AI Coding Revenue May Already Be a Fifth of the Entire US Developer Wage Bill
This is the question the group kept circling and never resolved. The BLS numbers are smaller than the industry assumes: 1.8 million US developers, maybe 200,000 at software companies and 600,000 at tech companies, for roughly $250 billion in total wages. If a large share of the labs’ enterprise API revenue is coding, they may already be near 20% of that entire market, and there have been none of the layoffs that a 20% productivity capture should produce.
Rory was deliberately careful not to draw a conclusion, and that restraint is the right call. There’s no hard evidence of a ceiling, and the momentum points the other way. But the arithmetic is worth holding: even a decelerated 6x next year off a $50-60 billion base is $300 billion, which is larger than the entire US developer wage bill. Jason’s floor is the physical one. You cannot spend more than you take in, even at 100% gross margins. Either coding revenue decelerates harder than anything in history, or a wall of CFOs is about to ask why tech spend doubled while headcount didn’t move.
10. The Bigger, More Durable TAM Is a 10% Software Tax on Everything
Rory’s second math is the larger, calmer one, and it’s the reason the two of them didn’t leave the segment purely bearish. Software spend is about $1.4 trillion and going agentic fast. If companies tolerate a 10% gross-margin spend on AI, the way they swallowed a 7% Amazon tax a decade ago, that’s another $100-140 billion in reach.
Jason’s proof, from running his own sales and marketing agent, is that outside of coding, agentic token costs are manageable, closer to 10% of revenue than 50%. Salesforce could comfortably pay 10% for tokens. It will never pay 40%, because it only runs 22% operating margins. So the honest map is three buckets: coding, the 10% agentic tax on all software, and eventually a co-work layer that goes after the knowledge worker at Office-like prices. The developer-wage ceiling is real, but it’s not the only source of demand.
11. SK Hynix’s $26.5B Listing Shows How Cyclical the AI Boom Still Is
Memory is where the capex boom shows up as a public-market cycle. SK Hynix priced the largest-ever NASDAQ listing by a foreign company, popped 13%, and gave it back, and the three memory players have ridden AI capex to operating margins that swung from negative in 2023 to 70% now. Rory’s bull case: they trade at 5 to 8 P/Es. His bear case: memory has always been capital-cyclical, and 70% net margins do not survive the next wave of capacity.
Jason’s connection is downstream. IBM’s stock crashed 20% this week and blamed memory directly: CIOs are scrambling to buy it before prices climb, and there’s nothing left in the budget for mainframes. It’s partly an excuse. It’s also a preview of the crowding-out to come, because when memory and tokens both bid up the same finite tech budget, something on the purchase order gets cut, and it won’t be the memory.
12. Late-Stage Venture Is a New Asset Class, Which Is Why Calacanis Going All-Growth Makes Sense
Calacanis moving his syndicate entirely to later-stage growth reads as a real marker, not a tweak, precisely because of how much he’d invested in the early-stage machine. The structural case for the move is strong: early-stage venture hasn’t grown much in two decades, but a roughly 5x larger late-stage business has appeared as companies stay private and go from zero to billions in five years. That’s a genuinely new category, nine-figure checks into companies already worth a billion, doing the job the NASDAQ small-cap market used to do. Deepening secondary liquidity, which Harry says he’s never seen this good, makes now structurally better than prior cycles rather than just frothier.
The caveat is comparative advantage, and it’s the part most people skip. Being a great first-check investor, the David Frankle skill, does not automatically monetize at $20 billion pre. Late-stage is a valuation-risk business, not a picking business, so the pivot is logical for someone optimizing dollars deployed and dangerous for someone whose real edge is being early. The category is new and it’s here to stay. That doesn’t mean everyone should reposition into it.
13. Constellation Buying TouchBistro at 1x Shows What Stalled Pre-AI B2B Is Worth
Constellation bought TouchBistro, a $70 million ARR Toast competitor, for $70 million. One times revenue. It’s a rare unobstructed look at what a stalled pre-AI B2B company is actually worth, with nothing in the way: no egos, no 2021 markup to defend, no founder drama. Venture debt from Francisco Partners had already converted to senior equity when the company missed and wiped out common, so the deal was just a spreadsheet.
Two lessons sit inside it. The debt one is loud and Jason is unequivocal: never stack a wall of debt on a slow-growth business, and never take debt in place of an equity round, because the misaligned cap table traps you. The scarier one is speed. In the agentic world these companies can’t be turned around, and the decay is accelerating past what the rollup math assumes. Marketo is the case study, going to zero with no net new logos, and Salesforce’s LLM-powered migration collapsed the switching cost from a year to a single day. When the switching cost goes to zero, terminal decay stops being slow.
Quotable Moments
Harry Stebbings
“Isomorphic was oversubscribed 8 to 10x. The majority of people who wanted huge amounts got nothing at all. That creates a blast radius, where suddenly there’s a lot of interest in Chai, and then a lot of interest in Latent Labs in London. It’s a blast radius from one very hot deal.”
“It really depends who says it. When I say that, I get in so much trouble. And then Paul Graham and YC say it and it’s ‘Yeah, marvelous.’ No one’s scared of me, that’s the difference.”
“The rise of tranched rounds isn’t new either. It’s just exacerbated by AI. These things aren’t brand new, they’ve just become normal for potential outliers, and there are just more potential outliers now.”
Jason Lemkin
“It’s 10 at night, and instead of doom scrolling, I’m doom coding. I go all night, and in the morning I want to check in on that workflow. I don’t know any AI developer who couldn’t consume even more frontier tokens. We just want more.”
“Marketo is going to zero. There are no net new logos, everyone hates it, and they just raised prices on us 20% again while deprecating the API. Salesforce did an LLM lift and it took one day to leave. It used to take a year.”
“If you don’t want to be worth 1x, do something before it’s too late, man. A slow-growth company at 6-8% with a big wall of debt on top of it? That’s your future, and it’s a brutal lesson.”
Rory O’Driscoll
“In the early stage, when you’re on the board, you can’t invest in two competitors. In the late stage, structurally, you have to. Fidelity Growth would have invested in both OpenAI and Anthropic. They’re not on the board of either. They want to make the secular bet.”
“It may well be that the prize for becoming the fastest-growing company in human history is that you hit TAM faster than any company in recorded history. You may have hit the limits of how much money there is in the till.”
“Treason doth never prosper. What’s the reason? For if it prosper, none dare call it treason. When you succeed as Uber, you get what you did legalized. When you succeed as Airbnb, you get what you did legalized. Cookie stuffing is a little less likely to clear that bar.”
This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll. The My Takes companion posts Sunday.
