cursor's warchest, xai's redemption

last month, cursor was closing a $50b oversubscribed round. the same week, the founders signed a deal selling the company to xai at $60b.

last month, cursor was closing a $50b oversubscribed round. the same week, the founders signed a deal selling the company to xai at $60b.

most fastest-growing-software-company-in-history founders don’t sell. cursor’s did.

the consensus read is that this is a great deal. xai needs an application surface to put in front of public market investors before the spacex ipo. cursor needs a sponsor with compute and a non-competing model lab. founders walk with billion-dollar exits. board books mark to $60b. customers don’t notice anything change. 1+1=3. genuinely a great deal for everyone in the room.

what’s worth sitting with is what the deal says.

cursor was the fastest growing software business in history. $2b arr in 13 months. 70% of the fortune 1000. enterprise stickiness, multi-year procurement contracts, a sales motion that beat copilot in head-to-head bake-offs. they did not need to sell.

their founders, who are extremely smart, looked at the path to $100b independent and decided they were not willing to underwrite it.

that’s the signal.

anthropic doesn’t tolerate middlemen

somewhere inside anthropic’s office there is a dashboard. it shows tokens consumed through cursor, tokens consumed through windsurf, tokens consumed through cline, tokens consumed through any of the dozen other harnesses sitting between anthropic’s model weights and the developer typing in vs code. and then it shows tokens consumed directly through claude code.

the kpi of every product manager looking at that dashboard is to make the second number bigger and the first number smaller.

every model lab whose customers are reachable through someone else’s distribution ends up here. the resellers will pricegouge as they scale, route to the cheapest model the day a competitor undercuts you, and own the user relationship that lets them swap you out at any time. anthropic watched cursor become the largest single buyer of frontier coding tokens on earth and correctly identified that as a problem rather than a partnership.

so they built claude code, and they priced it to win.

every token cursor sells passes through anthropic’s api at api rates. claude code is priced such that the effective per-token cost to a developer is roughly 5x lower than what cursor pays anthropic to resell. anthropic isn’t being nice in claude code. they’re running first-party agent margins close to zero specifically to keep the resale market underwater.

this is the same warchest dynamic openai burns the boats was about, run from the other direction. once a lab decides a profit pool is theirs, they burn the surrounding economics down until the people sitting on it leave on their own. openai did it to anthropic on api inference. anthropic did it to cursor on agentic coding.

and this is not a one-time campaign. every move anthropic has shipped in the agent space follows the same logic: pull tokens off third-party rails, route them through first-party rails, eat the margin hit short term, recover it once the resellers fold. claude code shipped, then claude code rate limits expanded, then claude code got cheaper plans, then anthropic started cracking down on third-party tools that resold or mirrored claude code's behavior. the playbook isn't subtle. anthropic has been telegraphing for a year that anyone disintermediating them was going to get squeezed. cursor was just the loudest test case.

for a year that anyone disintermediating them was going to get squeezed. cursor was just the loudest test case.

cursor tried everything

cursor was not lazy. cursor did not get outflanked because they didn’t see it coming. they ran every play in the book, and several plays that weren’t.

• composer. in-house model designed to absorb token-heavy workloads at a cheaper unit cost. it works. it just isn’t frontier, so the experience for hard problems still routes back to claude.

• agent harnesses for long-running tasks. cursor’s loop is more sophisticated than what ships in claude code, engineered specifically for runs that would be cost-prohibitive for anthropic to support directly. real product differentiation, real moat.

• business-user workflows. an attempt to escape the developer buyer entirely and sell coding agents as automation surface. either didn’t take off or didn’t move gross margins. probably both.

• lovable-style design features. ui generation, app builder surfaces, share-grabs in adjacent markets where anthropic wasn’t directly competing. same outcome.

• aggressive enterprise sales. cursor was the first ai app to seriously hit fortune 500 procurement cycles. they built sales teams while cognition was still a research lab. the timing lead is real and durable.

• open source models as a fallback. still not frontier on hard coding tasks. the gap is narrowing but not closed.

most of the discourse around cursor’s economics treats the negative 23% gross margins like they were the original sin of the business. they weren’t. sources close to the company say cursor was slightly gross-margin positive for at least the first year of the api-resale era, possibly longer. the negative 23% is a recent development. it is what the unit economics looked like after claude code spent a year aggressively underpricing the api, after gpt-5 dropped at $10 per million, after the entire pricing surface that cursor’s business was built on collapsed under pressure from its own suppliers.

the takeaway worth holding onto: cursor’s negative margins are the byproduct of a market that constricted on them faster than any sales motion could outrun, not a founder team that couldn’t read a cogs line.

so cursor’s options compressed:

1. burn until anthropic blinks. anthropic was not blinking.

2. train a real frontier model. composer is good, not frontier, gap is widening.

3. ride open source. still not enough.

4. find a sponsor with their own compute and a non-competing model lab.

option four was xai. compute, grok, no first-party coding agent at scale, very expensive ipo to support, cash. the deal terms got generous because the strategic fit was unique.

cursor took the offramp. it was the right move.

anthropic killed application and grew a peer

if you read the deal as a clean win for anthropic, you are reading it wrong.

cursor inside cursor was a margin-extracting reseller that swapped model providers in 24 hours. when openai dropped gpt-5, cursor was a launch partner before the press release was out. that overnight model swap is exactly what made cursor disposable, and it’s the same thing that made cursor strategically harmless: zero switching cost between model providers means you can be replaced too.

cursor inside xai is a different animal.

before this deal, xai was basically done — all their research talent had left, their models aren’t being used, the founders all quit. they had no enterprise contracts. they had no real developer user base at all.

now they own owns the largest enterprise coding distribution surface on earth, with a battle-hardened sales motion, deployment muscle that anthropic does not have, and a multi-year head start on procurement relationships at the fortune 500. they bolt that onto a frontier-model lab with one of the largest gpu fleets on the planet and a founder willing to lose money for a decade if the prize is right.

the engineers building cursor’s harness, the ones who spent two years figuring out how to make agents reliable at long-horizon tasks, are now grok’s engineers. anthropic took the most experienced applied coding agent team outside its own walls and handed them a bankroll. whether cursor-on-grok works as a product is genuinely unknown. there will be six to twelve quiet months of composer-on-grok-infra, xai bidding for fortune 500 contracts with the cursor sales team and elon’s rolodex, grok-coder benchmarks marketed as cursor launch material. nobody knows yet if the math works. but the strategic story is clear: anthropic just consolidated its frontier-model competition.

it also signals something to the rest of the market that’s worth chewing on. the founding thesis of the application layer was that you could build a neutral harness, pick the best model on a given day, and stay above the model fight. cursor was the textbook execution of that thesis, and the textbook execution couldn’t survive the squeeze. neutrality is a feature you can offer until your suppliers figure out you’re a margin to extract. then it stops working. or, more charitably, it stops working at the scale cursor was operating at. there is probably still a layer where neutrality survives. it’s just smaller than $50b.

what this means for everyone else

if you’re cognition, factory, lovable, replit, you just watched the most successful version of your business model decide it could not run independently. that does two things to your strategic position.

first, it lowers the bar for taking a sponsor. cognition does not need to wait until they’re bleeding 23 points of gross margin to entertain the conversation that cursor entertained. meta has llama, $50b a year of free cash flow, and zero credibility as an enterprise software company. amazon has bedrock and nothing on top of it that a developer would willingly use. microsoft has copilot and an openai exposure they’re trying to dilute. google has gemini and jules and nothing any enterprise has heard of. the call from at least one of these probably already came. it’s a question of timing.

second, it changes what you’re optimizing for. winning the category is no longer the only path that matters. doordash didn’t kill grubhub. doordash got to a scale where pursuing the last 15% of the market cost more than letting grubhub keep it. there is a niche-down strategy for the second-tier coding harness that involves picking a vertical, picking a buyer profile, picking a workflow, and being unkillable at that specific seam rather than fighting for the open seat at the fortune 500 procurement table. lovable already does this for design-heavy frontends. replit already does this for education and indie devs. that’s a real business. it just isn’t a $60b business.

third, for investors in the space, it means that investing in teams of genius viet-cong applied ai builders, fighting asymmetrical warfare against the big ai labs, can still pay off phenomenal returns, even if the vietcong don’t win the war. the market value of their distribution & R&D discipline can be worth $50B — much higher than any prior “acquihire” team (character.ai noam shazeer, adept, windsurf, etc.). this dramatically raises the ceiling on how easy it is to underwrite application layer companies even w/o a path to profitability

anthropic, post-victory

with cursor neutralized, the moves available to anthropic are obvious, and they’re already starting to ship.

claude code rate limits got tightened twice this quarter. third-party harnesses that mirror or wrap claude code behavior are getting api-banned in waves, including ones that were tolerated for a year and a half. tools that were previously living in the gray zone are getting cracked down on, not because they got worse, but because the strategic motivation for tolerating them is gone. with the largest disinter mediator of anthropic tokens on its way out — anthropic no longer needs to subsidize the ecosystem.

and none-too-soon either, as this happens alongside rumors of compute constraints, a push towards freeing up tokens for cowork, and

the warchest doesn’t need to stay full anymore. nobody’s parasiting on anthropic at scale, so the price has to start drifting somewhere closer to what the actual cost-to-serve plus a profit margin looks like.

uber spent a decade losing money to drive competitors out, then quietly raised rideshare prices 40%. claude code prices are about to start doing the same drift. the model serving the user gets better, sure, and also costs more, and also has rate limits that didn’t exist last quarter, and also no longer plays nice with the third-party harnesses that used to extract margin out of it.

the constraint is openai. anthropic can’t fully extract because gpt-5 still exists at $10 per million, gemini is still cheap, and the labs are still in a price war for distribution. but the cursor-shaped reason to subsidize claude code is gone. expect the price-quality curve on claude code to shift up over the next four quarters and the rate-limit envelope to tighten, in lockstep.

the read

cursor didn’t fail. cursor is the most operationally successful software company of the ai era by every metric saas investors care about. you cannot disintermediate the lab whose tokens you resell if they determine they want to go to war with you.

the lesson for everyone else watching is short. the application layer of ai doesn’t get champions. it gets wards. and every ward at scale ends up with a sponsor whose name is on the cap table when you decide independence isn’t underwritable.

thank you Jeff Wang, Gokul Rajaram, Russell Pekala, and Mark Hay for reviewing this essay