The Code Delusion

In a YC W26 demo day conference room, a 23-year-old founder showed a $2.4M ARR product. He had three customers in regulated insurance, one investor on the cap table, and zero engineering hires. He had written almost none of the code himself. Across the bay, a Series B SaaS company with 84 engineers was failing to ship a feature its customers had asked for nine months earlier.

The founder had no moat in his code. He knew it. His investors knew it. His pitch was about insurance carrier API access, a state-by-state compliance posture, and a contract with a broker network that would take four years for anyone else to replicate.

The 84-engineer company was still being marked at a 2021 multiple.

That gap — between what the market still pays for and what actually compounds — is the deception this memo is about. It is not about AI coding tools. Tools are the catalyst, not the story. The story is that the entire software industry built its pricing, its hiring, its equity grants, and its strategy around a proxy that broke in 2024, and most of the industry has not noticed yet.

Code is no longer the bottleneck. Code is no longer a moat. Code is no longer the deliverable. The cap tables, the org charts, and the term sheets have not adjusted.

This is what we now spend our days at Turing trying to evaluate. This memo is what we wish more LPs, founders, and operators understood.

The proxy that broke

For fifteen years, "lines of code shipped per quarter" was a workable proxy for software value. Engineering velocity meant product velocity. Product velocity meant moat. Moat meant valuation. The proxy worked because writing code was hard, slow, and expensive. The hard, slow, and expensive part is what got priced.

That proxy is dead. The death certificate is in the data.

Satya Nadella told a public audience at LlamaCon in April 2025 that "20%, 30% of the code that is inside of our repos today and some of our projects are probably all written by software." (CNBC) GitHub's 2025 Octoverse reported that Copilot generates an average of 46% of the code its users ship, with Java reaching 61%, and that 80% of new developers on GitHub use Copilot in their first week. (GitHub) Stack Overflow's 2025 Developer Survey put AI tool usage at 84% of developers, with 51% of professionals using AI tools daily. (Stack Overflow)

The economic translation is starker than the percentages. Anthropic's Claude Code went from launch to $2.5B in annualized revenue between February 2025 and February 2026. (SaaStr) Cursor crossed $2B ARR in three years. (The Next Web) Replit's ARR went from roughly $10M at the end of 2024 to $240M for full-year 2025. (Inc.) These are not niche tools. They are infrastructure. And they have collapsed the unit cost of producing functional software by something like an order of magnitude.

The industry's cap tables, hiring plans, and strategic narratives have not adjusted to a world where the marginal cost of code went to zero.

What the data says vs what the cap table says

If code were genuinely cheaper, software multiples would compress for code-only businesses and expand for businesses that own something a model cannot generate. The first half of that prediction is happening. The second half is being faked.

The Bessemer Cloud Index's median public SaaS multiple sits around 7.5x EV/Revenue as of early 2025, down from a September 2021 peak of 18.43x. (Bessemer) That is a 60% compression in the multiple paid for "we sell software" without any other story. Carta's State of Private Markets data shows Series A median post-money at $78.7M and seed median post-money at $24M in late 2025 — record nominal highs, but the entire spread is being driven by AI labels. AI companies now capture 42% of seed capital, up from 23% pre-ChatGPT, and command a 38% Series A valuation premium over non-AI peers. (Carta)

That premium is being paid because investors think they are buying scarcity. They are not. They are buying companies whose primary input — code — has become the commodity. Crunchbase reports that roughly 50% of all 2025 venture funding went into AI-related companies, with $211B raised, up 85% year-over-year. (Crunchbase) Most of that capital is funding a category whose marginal cost is collapsing while it is being priced as if it were rising.

Read the public-private spread. A median AI Series A is being marked at ~10x revenue or more. Public SaaS sits at 7.5x. Either the public market is wrong, or a generation of seed and Series A rounds will compress 40–50% before they hit liquidity. We think the public market is right and the private market is processing the news slowly.

The hiring fiction

The labor market is the clearest place to see the lie.

Layoffs.fyi and Crunchbase put 2025 US tech layoffs at roughly 127,000 workers, with about 69,840 of those job cuts directly linked to AI restructuring. (Crunchbase) Software engineers were the single most-cut role in those restructurings. Salesforce announced a hiring freeze on engineers as it credited AI for productivity gains. (IT Pro) Mark Zuckerberg told Joe Rogan in early 2025 that Meta would have AI doing the work of mid-level engineers within the year, and that mid-six-figure salaries for that tier would not survive. (Entrepreneur)

So far, so consistent with the thesis.

Now look at what the same companies actually did. Stripe cut 300 engineering and product roles in early 2025 and immediately announced it was raising headcount to 10,000 by year-end. (Irish Times) Carta's compensation data shows new-hire engineering and product salaries at $189K averages, with engineering equity grants only ~15% smaller than a year earlier and overall equity ~26% below pre-2022 — meaningful, but not the collapse the productivity story would predict. (Carta) Klarna, the most public AI-replaces-humans case study, walked the entire bet back: Sebastian Siemiatkowski admitted in 2025 that the company's aggressive replacement of 700 service workers had "gone too far," and Klarna began rehiring through 2025 and 2026. (Entrepreneur)

The industry is caught between two contradictory lies. The first lie says engineers are still scarce, so we should keep paying them as if 2021 multiples were going to come back. The second lie says AI has replaced engineers, so we should fire them and post the savings to operating margin. Neither is true.

The truth is uglier. Most engineering work in commodity SaaS has been quietly commoditized — scaffolding, CRUD, glue code, integrations. The senior engineers who can architect a system, design data infrastructure that compounds, navigate regulated workflows, or ship a model into a production environment without breaking it are not commoditized, and their leverage has gone up. The mass middle has gotten cheaper and the edges have gotten more expensive. Companies are pricing the average and getting both ends wrong.

The productivity paradox nobody wants to discuss

If AI coding tools were the unalloyed productivity story their vendors sell, organizational metrics would show it. They do not.

The METR study published in July 2025 took 16 experienced open-source developers and randomized them into AI-allowed and AI-disallowed conditions on real tasks in mature codebases. Developers predicted a 24% speedup. They self-reported a 20% speedup after the fact. The actual measured outcome: AI assistance made these developers 19% slower. (METR) Time saved on writing was eaten by reviewing AI output, prompting, and waiting.

GitClear's analysis of 211 million changed lines from Google, Microsoft, Meta and enterprise C-Corps showed a different fingerprint of decay. Refactored code dropped from 24.1% of changed lines in 2020 to 9.5% in 2024. Copy-pasted code rose from 8.3% to 12.3%. Code revised within two weeks of being written rose from 5.5% to 7.9%. (GitClear) 2024 was the first year on record where new copy-paste exceeded refactoring. The codebases are getting more derivative and more brittle.

Snyk's 2025 AI Code Security report found that 56.4% of developers frequently encounter security issues in AI-generated code, while 80% bypass their organization's AI code security policies anyway and 75% believe AI-written code is more secure than human-written code. (Snyk) Bain's 2025 software development report found that teams using AI coding assistants saw 10–15% productivity boosts, but the time saved rarely got redirected into higher-value work, so even those gains failed to convert into business returns. (Bain)

Google's own 2025 DORA report nailed the contradiction. AI adoption is at 90% across surveyed organizations, individual contributors complete 21% more tasks and merge 98% more pull requests — and organizational delivery throughput stayed flat. (DORA) The output went up. The outcome did not.

This is what a productivity proxy looks like when it has detached from value. Lines shipped is up. Stuff that matters is roughly the same. The industry is paying for the lines.

Where the moat moved

If code is not the moat, what is? The companies that have raised at the highest defensible valuations in the past twelve months tell a remarkably consistent story. None of them are selling code.

Harvey crossed $190M ARR by January 2026 and raised at an $11B valuation in March 2026. (Bloomberg) Its moat is not its model. Its moat is that 50 of the AmLaw 100 firms have built workflow on top of it, and a competitor would have to break legal liability assumptions partner-by-partner to displace it.

Abridge hit $100M ARR by mid-2025 and raised at a $5.3B valuation, then again at higher in early 2026. (Sacra) Its moat is a proprietary corpus of 1.5 million medical encounters and an Epic distribution deal — equity stake plus revenue share — that gives it preferential access to 38% of US hospital networks. (Contrary Research) Anyone could copy the code in a quarter. Nobody can copy the data or the Epic relationship.

OpenEvidence went from $3.5B to $12B in valuation between July 2025 and January 2026. (STAT) Its asset is not its model — its asset is being the AI-medical-reference product physicians have already integrated into clinical workflow at scale.

Sierra reached $100M ARR in seven quarters, raised at a $10B valuation in September 2025, and prices on outcomes — pay-per-resolved-conversation — not seats. (TechCrunch) Bret Taylor has been explicit that distribution to enterprise customer service organizations and the integration depth that comes with outcome-based pricing is the moat, not the agent.

EvenUp, the personal-injury legal AI, raised $150M Series E at over $2B in October 2025 on the back of a proprietary dataset of hundreds of thousands of injury cases and millions of medical records. (Fortune) The model is replicable. The dataset is not.

Tennr raised $101M Series C at a $605M valuation on a vision-language model trained specifically on medical referrals against payor criteria — narrow data that "wouldn't make financial sense for broader competitors to pursue," in the founders' framing. (Fortune) That is regulatory-and-data depth as moat, not architecture as moat.

Figure raised at $39.5B for a humanoid robot that has done 1,250 hours and 90,000 sheet-metal placements at a BMW plant, and Physical Intelligence added $400M for foundation models for physical world manipulation. (The Robot Report) The bet there is hardware-software integration so deep it cannot be reproduced in software alone.

Read the names again. The pattern is uniform. Every large 2025 AI win was sold as a "model company" in the press and is actually a distribution, data, or regulatory company in its cap table economics. The model is the wrapper. The moat is what surrounds it.

The YC tell

The clearest leading indicator of where the industry is going is what the next generation of founders is choosing to build, who is building it, and how fast they are getting to revenue.

Y Combinator's W26 batch was 199 companies. Twenty hardware startups. Three AGI labs. Twenty-two solo founders — 11% of the batch — building alone. (Extruct) Three times as many companies in W26 reached $1M annualized revenue as in W25. The median AI agent founder in the batch had 4.8 years of professional experience. The median time to revenue for batch companies has compressed dramatically.

Garry Tan has been explicit in interviews that the new YC playbook involves AI tools doing the work that founding engineers used to do, and that the unicorns of the next five years will not look like the unicorns of 2015. (Vanta) Sam Altman has spent the last two years openly speculating about when the first one-person, one-billion-dollar company arrives, with a betting pool among CEO friends on the year. (Fortune)

The implication is uncomfortable for the rest of the industry. If a 22-year-old solo operator can ship a $1M ARR product in W26 with no engineering team, then the "founding engineer" market — the equity bidding war over the second hire at a Series A — is being priced for a world that already ended. Carta data shows the first hire still receives ~1.5% equity median, with engineering equity grants down ~15% from a year ago but still substantial. (Carta) That is not a market in which scarcity has been priced out. It is a market that has not yet finished translating the news.

The W26 batch is the canary. The S26 and W27 batches will tell us whether the canary is dead.

The counterargument and where it actually holds

The thesis breaks in three places. We want to be honest about them, because they matter for where Turing actually deploys capital.

First, foundation model companies. OpenAI, Anthropic, xAI, and a handful of frontier labs are still selling code — model weights, training infrastructure, inference systems. Anthropic just hit $44B ARR as of May 2026. (SaaStr) The labs are the exception that proves the rule, because their code is the input that commoditized everyone else's code.

Second, infrastructure with performance constraints. Custom silicon (Cerebras, Groq), database internals (the Snowflake / Databricks / DuckDB layer), high-frequency systems, real-time inference orchestration. These are domains where the code itself encodes years of engineering taste that AI cannot yet replicate, and where the cost of being wrong by 5% on latency is the difference between a product and a science fair project. We think this is roughly the bottom 5% of total engineering employment but a much larger fraction of strategic value.

Third, where the data the model has not seen is what the customer is paying for. Cognition (Devin) is in conversations to raise at a $25B valuation (Bloomberg), Replit at $9B, Cursor at $50B in talks. These companies are selling code to engineers who use less code from each other. The fact that this category is the fastest-growing software category of 2025 is the most expensive irony in the industry.

But carve out those three exceptions, and the thesis holds across what we estimate is more than 90% of total private software valuation. The horizontal SaaS application company that won 2014–2021 — sells software, charges per seat, beats competitors on velocity — is in a bad spot. The vertical, regulated, data-rich, distribution-deep operator company is in a great one.

Simon Willison, the most thoughtful operator-essayist on AI-assisted programming, has been explicit that "vibe coding your way to a production codebase is clearly risky" and that experienced engineers reviewing every line still produce the durable systems. (Simon Willison) Karpathy, equally explicit, calls the current moment "summoning ghosts" — non-deterministic systems whose behavior we are bolting onto deterministic infrastructure. (karpathy.bearblog.dev) Both are right. Neither contradicts the thesis. The fact that AI-generated code is fragile is what makes the data-and-distribution moat more valuable, not less. The competitor's code being shaky does not save you. Your having a moat that is not code does.

What investors should actually evaluate

The Coatue and First Round and Khosla research notes are full of frameworks. We will offer one, calibrated for early-stage AI software in 2026.

When we evaluate a company at Turing, we now ask six questions before any question about engineering velocity:

1. Where does the data come from, and who else can get it? Proprietary corpora and access agreements (Abridge–Epic, EvenUp's case archive, Tennr's referral documents) are durable. Public-internet-scraped data is not.
2. What is the regulatory or liability surface, and how deeply is the company entangled with it? Harvey's relationship with AmLaw 100 firms, Tennr's payor criteria modeling, Sierra's enterprise compliance posture — these slow down replication far more than software architecture does.
3. How does the company price? Outcome-based pricing (Sierra) and revenue-share pricing (Abridge) bind the company to customer success in a way per-seat SaaS pricing does not. Per-seat SaaS pricing is, at this point, a leading indicator of disruption of the seller.
4. What is the founder's distribution motion? A 24-year-old who has shipped to insurance carriers and brokers in three states is more interesting than a 32-year-old who has shipped a beautiful internal tool. The bottleneck has moved from build-it to land-it.
5. What does the equity grant for the second technical hire look like? Companies still grafting 1.5% to a "founding engineer" are not telling investors something useful — they are telling them they have not internalized that engineering is no longer the rate-limiting constraint.
6. If the company's primary model were leaked tomorrow, what fraction of revenue would survive ninety days? That is the moat number. It is the only one we trust.

Patrick Collison framed this in early 2025 with the right metaphor: software, he said, "should be like pizza, cooked right then and there at the moment of use." (a16z) The fixed-cost-then-infinitely-monetized model that defined 2014–2021 SaaS is being replaced by one in which inference is variable cost and customer ownership is the only durable input. The companies that internalize this and the ones that fight it will diverge sharply in the next 24 months.

The close

The deception in the title of this memo is not malicious. It is just lagging. Most industries take a decade to reprice after a structural input collapses. Software has had eighteen months. The cap tables, term sheets, equity grants, and headcount plans of the median 2026 software company are still constructed as if the labor required to produce code were the bottleneck of the business. It is not, and the gap between the cap-table assumption and the underlying economic reality is what the next several years of returns will trade against.

The companies that win the next decade will be the ones that admit, internally and pricelist-publicly, that code is no longer the deliverable. They will price on outcomes. They will fight for proprietary data. They will marry themselves to regulatory regimes that scare off generalist competitors. They will hire fewer engineers and more operators, lawyers, salespeople, and domain experts. They will treat their model as a wrapper around a moat, not as the moat itself.

The companies that lose the next decade will keep paying 1.5% to founding engineers, hiring against a lines-of-code roadmap, and pitching velocity decks to investors who have stopped pricing velocity.

Most of what gets called software in 2026 will not exist in 2030. The market is already pricing this in the public comps. The private market will catch up. The companies that catch up first will look mispriced today and obvious in retrospect.

That is the trade. It is the trade Turing is built to make.