A founder raises $125M, hits $75M ARR faster than any SaaS company in history, and then one Tuesday morning the capability ships natively inside ChatGPT. The product becomes a feature. The Series B money is still in the bank. The thirty-five engineers are still on payroll. The company is already dead. It just has eighteen months of runway before anyone else notices.
This has happened approximately four thousand times in three years.
The body count is large enough that the industry has started to joke about it, though the jokes have the brittle quality of people laughing at something they are also afraid of. I want to write the history of this graveyard as a diagnostic, not a roast. The pattern has a specific geometry, and the geometry tells you exactly what is about to happen to the next wave of startups now that GPT-5.5 can code autonomously for twenty hours and Claude Mythos can do things Anthropic won't let it do in public.
The graveyard is not done filling up. In some ways, it has barely started.
012023: the Goldrush
The starter pistol fired on November 30, 2022, when ChatGPT launched. Within six months, every venture firm in Silicon Valley had rewritten its thesis to be "AI" and every founder had rewritten their pitch deck to include "GPT-4 integration." The 2023 vintage of AI startups was, in retrospect, the most naive cohort in tech history. Not because the founders were dumb, but because the market structure they were betting on did not yet exist and would not be visible for another eighteen months.
Jasper raised $125M at a $1.5B valuation on a simple thesis: wrap GPT in a marketing UI, sell to SMB content teams. Peak $75M ARR in under two years, faster than any SaaS company in history. The category exploded in its wake: Copy.ai ($80M raised), Writesonic, Rytr, Anyword, dozens more. Pure Aggregation Theory on autopilot. Distribution beats underlying tech, own the customer, ride the capability curve.
Character.AI: $150M from a16z at a $1B+ valuation for a consumer AI companion product with hundreds of millions of users. Inflection AI: $1.3B to build Pi, a "kind, curious, helpful" consumer assistant; Reid Hoffman and Mustafa Suleyman co-founding; treated as a third foundation-model lab alongside OpenAI and Anthropic. Adept AI: $415M for an agent that could "take actions on the web." Tome: $43M at $600M to be AI-native PowerPoint. Humane: $230M for a wearable pin to replace the smartphone. Founders on Twitter declared the iPhone dead.
Meanwhile, the substrates (OpenAI and Anthropic) were charging ahead. Context windows from 8K to 100K to 200K. Models smarter, cheaper, faster, on a cadence that surprised even the labs.
Nobody noticed the graveyard was already being built. The shovels were arriving. Nobody had died yet.
02The Pincer, Named
Here is the geometry that makes sense of everything after.
An AI startup in 2023–2026 operates in a narrow strip of market territory. Above it sit the model companies (OpenAI, Anthropic, Google DeepMind), constantly absorbing features from the ecosystem into their first-party products. Below it sit the incumbent platforms (Microsoft, Google, Salesforce, Adobe, Notion, the hyperscalers), constantly absorbing AI capabilities into their existing distribution surfaces. The startup is caught between them like a mouse on a kitchen counter with cats approaching from both ends.
This is the Pincer. Not a metaphor. The literal market structure that killed or is killing every company I am about to list.
The top pincer closes when a model company ships your feature natively. GPT-4 summarization kills standalone summarizers. ChatGPT Memory kills personalized-AI wrappers. Claude Projects kills knowledge-base-over-LLM products. Code Interpreter kills data-analysis startups.
The bottom pincer closes when an incumbent bolts AI onto their installed base. GitHub Copilot killed the standalone code-completion market; Tabnine and Kite were dead the moment every VS Code user got autocomplete for free. Salesforce launched Agentforce and told every enterprise buyer to cancel the AI point solutions they were evaluating. Notion AI went from paid add-on to free for every workspace, suffocating an entire category of note-taking-AI startups overnight.
And sometimes the pincers don't just squeeze the startups; they turn on each other. On April 14, 2026, Mike Krieger, Anthropic's CPO and co-founder of Instagram, quietly resigned from Figma's board. Three days later, Anthropic launched Claude Design, a standalone AI workspace that generates UI mockups, prototypes, and design systems from conversation, a direct competitor to the product Krieger had been helping govern. Figma stock dropped 7.5%. The company that had been integrating Anthropic's models as AI assistants for its own users discovered that its AI supplier had become its AI competitor. The top pincer doesn't just eat startups. Given enough capability, it eats the incumbents too.
The startup in the middle has two problems, not one. And the problems compound, because the better the startup's product, the more attractive the territory becomes to both attackers at once. You cannot escape the pincer by being good. Being good is what attracts the pincer.
Packaging value, in this geometry, is the dangerous act. A well-packaged capability is a well-marked target. The prettier your box, the faster the two bigger boxes on either side reach over and take what's inside.
032024: the Absorption Wave
2024 was when the pincer started closing.
Top pincer. January: OpenAI ships GPT Store and custom GPTs, absorbing every personalized-AI wrapper pitch in a single launch. February: Sora ships, wiping out AI video startups overnight. May: GPT-4o adds native voice. Hume, Play.ht, and a dozen others watch the ground disappear. October: Anthropic ships Computer Use, and every "AI does things on your computer" startup had their core capability demonstrated in a forty-second demo video.
Bottom pincer. Microsoft Copilot went from punchline to $10B+ run rate, embedded into every Microsoft 365 seat. Salesforce launched Agentforce and explicitly pitched it as "the replacement for AI point solutions you're currently evaluating." Notion AI went from $10/month add-on to free for every workspace. The incumbents had stopped being slow.
The bodies. March: Inflection, the $1.3B consumer AI company, had its entire team absorbed by Microsoft in a "licensing deal" that was functionally an acqui-hire and killed the company. Pi still technically exists. Nobody talks about Pi. June: Adept AI ($415M raised, $1B valuation) had its top engineers hired by Amazon. The idea wasn't wrong; Amazon wanted the team for exactly the idea's reason. The startup couldn't move fast enough to matter. August: Character.AI, with hundreds of millions of monthly users, talent-acqui'd by Google for $2.7B. The founders went back to Google, where they'd originally worked before quitting in frustration with Google's slowness. Valuation written down to near zero inside a year. November: Humane shipped the Ai Pin. MKBHD called it "the worst product I've ever reviewed." Sold for parts to HP; $230M not recovered.
The grownups in the room realized the pincer existed. The founders mostly didn't, because they were heads-down trying to hit Series C milestones, and the pattern was only visible from outside the specific fight you were losing.
042025: Consolidation and Confusion
2025 was when the pincer became common knowledge and the venture capital machine, in its infinite wisdom, poured a record $192.7B into AI startups anyway.
This was not purely irrational. Real winners were emerging, and they created narrative cover for the whole category. Cursor hit $500M ARR mid-2025 on its way to $2B. Midjourney was printing cash at ~50 employees, bootstrapped. Perplexity at $9B. Harvey signing top law firms. Glean hitting $200M ARR enterprise. These companies made it look like the application layer was alive and well, which it was, but only for a very specific subset that had accidentally escaped the pincer's geometry.
Meanwhile the pincer kept closing on everyone else.
OpenAI shipped o1 (September 2024) and o3 (January 2025), reasoning models that ate every "chain-of-thought wrapper" startup inside three months. Operator launched in January, a general web agent. Deep Research launched in February, consuming every AI-research-assistant startup.
By mid-2025, the top pincer had closed on essentially all of: generic AI writing, coding assistance, research, summarization, email, scheduling, voice transcription, and browsing. If your product could be described as "AI that does X," and X was a task a knowledge worker did on a computer, the top pincer had probably closed on you already.
More bodies. Builder.ai collapsed in May 2025 in a fraud scandal. $450M raised at $1.5B on "AI-assisted app building" that turned out to be human developers in India. Mem Labs wound down. Tome pivoted entirely out of AI, which is the corporate equivalent of a nuclear submarine filing for a fishing license. Jasper, by this point a zombie, rebranded as "agent workspace for marketing," a pitch with the desperate energy of a band announcing a new direction three albums too late. And in early 2026, Delve, an AI compliance startup valued at $300M, founded by MIT dropouts who made Forbes 30 Under 30, was expelled from Y Combinator after it emerged that their automated SOC 2 audits were, to put it gently, not audits. The company named after the single most ridiculed word in the ChatGPT lexicon turned out to be generating fake compliance reports. You cannot write satire faster than this industry generates source material.
End of 2025, the industry had a new parlor game: name an AI startup that raised above $50M and is more valuable today than at its peak. The list is short. Cursor, Midjourney, Perplexity, Harvey, Glean, Sierra. Maybe five or six more. Then it ends.
05Who Escaped, and Why
Before I get to the present (which is worse), the survivors deserve a pause, because their pattern is instructive.
The pincer eats everything it can digest. The companies that survived were the ones it couldn't. Not smarter, not faster, not better-funded. Inedible. Each had some structural property (a weird shape, a bitter taste, a shell too hard to crack) that made the cost of absorption higher than the value of the territory.
Cursor. The poster child, for a while. OpenAI and Anthropic were structurally uninterested in making the best IDE. Microsoft, which does make an IDE (VS Code), was in bed with OpenAI and structurally conflicted about launching a direct competitor. Cursor sat in the gap created by that conflict, widened by six-months-deep proprietary user workflow context. It was inedible because neither jaw wanted to become an IDE company. But inedible is not the same as immortal. This week, SpaceX, which absorbed xAI earlier this year, announced an option to acquire Cursor for $60 billion. The company that escaped the pincer is being swallowed by something else entirely: a vertically integrated AI empire that wants to own the coding surface from model weights to keystroke. Meanwhile, the developers on the cutting edge have already migrated to Claude Code, a terminal-native agent that doesn't need an IDE at all. Cursor may have escaped the pincer only to discover that the pincer was the wrong thing to be afraid of.
Midjourney. Never took venture capital, so never had to chase a valuation curve that would force it into pincer-adjacent territory. Accumulated a cult creative community and a distinctive aesthetic. Inedible to the bottom pincer because Adobe can't replicate taste. Inedible to the top pincer because model labs can replicate quality but not identity. The thing protecting Midjourney is not a moat in the traditional sense; it's that the product is the community, and communities can't be shipped in a blog post.
Harvey, Glean, Sierra. Vertical specialization, deep enterprise integration, outcomes-based pricing. Inedible because the top pincer has no interest in becoming a legal-tech or enterprise-search or customer-service company; those require armies of industry-specific integration work, messy sales cycles, and domain expertise that doesn't compress into a model release. The bottom pincer is too horizontally committed to vertical-specialize fast enough. These companies survived by being shaped like something neither jaw was designed to chew.
The lesson is not a playbook. The lesson is that survival in this market is a question of digestibility. The pincer will eat anything it can reach and swallow. The only companies still standing are the ones it bit into and spat out, because the effort of chewing exceeded the nutritional value of the territory.
The question for anyone building now is whether any inedible territories remain. The answer is yes, but they are harder to find, less glamorous to pitch, and require a kind of founder who knows where the hairballs are.
06April 2026: the Pincer Closes Fully
Which brings us to April 2026, the month the pincer finished closing.
April 8. Anthropic shipped Claude Mythos Preview. Not an upgrade to Opus, but an entirely new tier above it, codenamed Capybara, "larger and more intelligent than our Opus models, which were, until now, our most powerful." 93.9% on SWE-bench. 97.6% on USAMO. Found thousands of high-severity zero-day vulnerabilities in every major operating system during testing. Anthropic decided it was too dangerous to release publicly. Distributed through Project Glasswing, a gated preview to critical infrastructure operators only. The U.K. AI Security Institute confirmed it as the first AI model capable of completing their full-network-attack simulation.
April 23, today, as I write this. OpenAI shipped GPT-5.5. The first model trained from scratch since GPT-4.5. Not a fine-tune, a full rebuild. 82.7% on Terminal-Bench 2.0. Outperforms GPT-5.4 on Expert-SWE, OpenAI's internal eval for long-horizon coding tasks with a median human completion time of twenty hours. The model autonomously allocates reasoning time per task; complex refactors can run seven-plus continuous hours of autonomous work. NVIDIA has ten thousand employees on it. Jensen sent a company-wide email with the subject line "let's jump to lightspeed." 85% of OpenAI itself uses Codex every week.
Two launches, fifteen days apart. Both push the capability boundary well past anything an application-layer startup could plausibly differentiate against.
In pincer terms, simple and devastating.
The top pincer just got deeper teeth. A model that can autonomously complete twenty-hour engineering tasks absorbs the territory of every "AI for engineers" startup that wasn't named Cursor. Every founder whose thesis was "we wrap a model and do [knowledge-work task X]" now has to explain why a model that can autonomously do [task X plus everything around it for seven hours straight] wouldn't just do the whole thing itself inside ChatGPT or Codex. Most cannot explain it. Most founders raising seed rounds for "AI agents that do Y" in Q2 2026 are raising for companies whose moat was just obliterated in a blog post they probably haven't read yet.
The bottom pincer just got distribution. GPT-5.5 is embedded in Codex, which is rolling out to every enterprise with a ChatGPT Enterprise seat via SDK. The incumbent distribution players (Microsoft, Google, Salesforce) now have frontier-capable agents available natively in the surfaces where work already happens. The gap between "fancy AI startup with a better UI" and "Microsoft 365 seat with GPT-5.5 inside" just went from large to uncrossable.
And Mythos is the warning shot about what comes next. A model that finds zero-days autonomously today is a model that, with a few more capability generations, does arbitrary multi-step technical work autonomously. We are about twelve months from a capability surface where most knowledge-work verticals (legal research, clinical documentation, financial analysis, consulting deliverables) are performable autonomously by a frontier model with basic tool access.
The territory safe from the pincer is shrinking in real time, on a cadence measured in model release cycles.
07The Closer
I will resist ending with a tidy prescription, because I don't have one. The founders walking into this market deserve the truth more than a playbook.
The graveyard you can see right now (Jasper, Character.AI, Inflection, Adept, Humane, Rabbit, Tome, Mem, Builder.ai, Clara, Copy.ai) is the visible graveyard. The bodies that piled up when the pincer was still closing. Those companies had 18–24 months of warning before the jaws reached them.
The graveyard about to be built is different. GPT-5.5 shipped today. Mythos shipped two weeks ago. The cycle time from "capability released" to "startup obsoleted" has compressed from eighteen months to something closer to ninety days. The founders building right now will not have the luxury of pivoting three times on their way to extinction. They will get pivoted once, by a blog post they didn't see coming, and then the runway will do the rest.
The $192.7B deployed in 2025 will have mark-to-market returns that are, on aggregate, catastrophic. A handful of winners (the Cursors and Midjourneys and Harveys) will carry the returns. Most of the rest will be quietly written down over 2027–2028. The founders who survive the next cycle will not look like the ones who thrived in the last. The 2023 cohort optimized for speed-to-product on top of a capability platform. The 2026 cohort needs to be optimized for something much harder: finding a business structurally inaccessible to two trillion-dollar-scale attackers at once. That is not a founder profile that shows up in accelerator applications. It's closer to an industry veteran with a domain moat, someone who knows where the real data is and has the relationships to ingest it. The accelerator-grad-with-a-GPT-wrapper era is over. The industry-veteran-with-a-domain-moat era is starting.
Pack your value too neatly, and the pincer takes it. Pack it messily enough that neither side wants the territory, and you might live. The next wave of founders has to figure this out on a 90-day clock, with the jaws already visible on both sides of the room.
Most of them won't. The graveyard is going to be enormous. It is going to be beautiful, in its terrible way. The largest capital misallocation in the history of venture capital, happening in full view, in real time, on a cadence set by model release cycles that nobody in any boardroom controls.
Welcome to 2026.