Many won’t pass through the Slop Chasm

Technology works for us because it is reliable and predictable. This is so obvious that you barely think about it until it breaks. If your front-door key failed to turn one time in a hundred, you would buy a new lock that afternoon. A window that does not seal on an aircraft costs lives. If your electricity were not there 99.99% of the time, you would find a new provider or buy a generator. Keep this standard in mind.

The AI boom is real. Over time it will change industries the way electricity did: portable power let people live in denser, safer, richer cities, and portable intelligence will do the same.

The financial story is different. Geoffrey Moore called the step from early adopters to the pragmatic majority “crossing the chasm”.² The pragmatic majority buys technology that just works. AI is attempting that crossing right now by going public, and it is not ready. Most companies in the space will fall into what I call the Slop Chasm: the gap between a probabilistic demo and finished, usable work.

The claw machine shows the gap. A language model is probabilistic. In a ten-step job (a monthly close, a twenty-tab spreadsheet), a grip that holds 98% of the time delivers the prize 81.7% of the time.¹ One job in five falls back into the pile, and the coin stays in the box.

Intuitively, you already know the crossing has not happened.

1. If AI content is ready and so good, where are the paid AI streaming services with millions of subscribers?
2. Which everyday products are decisively better today because of AI?
3. Which non-tech companies have lapped their competitors on the bottom line because of AI?

A year is not a long time. Yet the valuations and the spending commitments force you to act as if all of this has already happened. NVIDIA, SpaceX and Anthropic together are worth roughly $9 trillion: NVIDIA about $5.5 trillion, SpaceX near $2 trillion, Anthropic at the $1.5–2 trillion its IPO is rumored to seek.³ The IMF puts the 2026 output of Germany and Japan, together, at $9.8 trillion.⁴

Economically, the market is telling you that you would rather own GPU clusters and satellites than sushi, beer or BMWs. The rest of this essay is about whether the clusters get paid.

The industry is complicated, but not that complicated. It is electricity for mental work, in two physical stages.

Generation is megawatts turned into silicon: chips, buildings, power. Once financed, the capacity is fixed. Anthropic’s filing says it plainly: compute is procured “12 to 18 months in advance of availability”, so the expense is “generally fixed in the near term”.⁵ Like an airline seat or an LNG cargo, an unused GPU-hour at takeoff is gone forever, and if you overshoot you cannot easily remove capacity.

Transmission is tokens per second: memory bandwidth, speculative decoding, quantization, caching. Engineers keep squeezing more tokens out of the same silicon; the model behind this essay assumes serving throughput doubles each year.⁶ That inflates effective supply without a single new chip.

Alphabet shows the economics of a grid at scale. Its first-party model APIs now process nearly 32 billion tokens a minute, about 530 million a second, four times a year ago.⁷ Assume, generously, that a third of those are paid at a blended $1.71 per million tokens. That is about $9.6 billion a year, or 45 cents per chip-hour across a 2.4-million-chip fleet.⁸ Take out 13 cents of power and a chip clears under $3,000 a year; a $30,000 chip does not earn its silicon back in six years. Alphabet can carry this because Google Services earned $39.2 billion of operating income last quarter and Cloud another $11.1 billion. A standalone lab has no such boiler room.

Revenue less cost of revenue implies a gross margin near 85%.¹¹ Look instead at Note 8 of Anthropic’s S-1, where expenses are shown by function for the first half of 2026.⁹

Cost of revenue, “substantially all” compute for customer-serving workloads plus some people and overhead, was $2.38 billion. Research and development was $9.74 billion: mostly compute for model development, research and internal use, plus people and software. Sales and marketing was $3.81 billion, and the filing says what sits inside it: channel partner costs, and “compute and infrastructure costs associated with providing free access” to Claude. Marketplace sales are booked gross, so the cloud partners’ cut lives here, not as a discount. Total operating expenses: $16.98 billion.

Cost of revenue is 14% of that. MBI Deep Dives, working from the same note, puts customer-serving compute at about 27% of the first-half compute bill, down from 44% a year earlier.¹⁰ These are dollar shares, not counts of GPU-hours, and the filing itself warns that cost of revenue includes personnel and overhead. But the direction is not in doubt: roughly three-quarters of what the company spends on compute serves research, model development and free users. The 85% describes an accounting line. The fleet is mostly a research fleet and a free fleet, and the bill for all of it is fixed.

When someone says an agent task “used 500,000 tokens”, the chart is counting every token the model read on every step. Most of that is the same context read again. Billing knows the difference. At Anthropic’s list prices for the Opus 5 generation a fresh input token costs $5 per million, an output token $25, and a cached re-read 50 cents, a tenth of input. On the newest models the cached rate is a twentieth.¹²

The mix is measured, not imagined. OpenRouter reported on October 8 that in September cached prompt tokens were just under 90% of all agentic tokens on its platform, up from about 60% in January, and that agent API keys used about seven times the tokens of human ones.⁴¹

Take an agentic task where 89% of tokens are cached context, 10% fresh prompt and 1% output and reasoning: 445,000 cached, 50,000 fresh, 5,000 output. The bill is 22 cents, 25 cents and 12.5 cents. Sixty cents. Price the whole 500,000 at $5 per million and you would have guessed $2.50, four times too much.¹³

Compute tells the same story. A cached token skips the expensive first read; an output token is generated one at a time and costs the most. In fresh-token equivalents the task is about 97,000, not 500,000: a fifth of what the chart implies. The one percent of output tokens is a quarter of the compute and a fifth of the money.

So the exploding token charts, on public routers and in Google’s 32 billion a minute, count usage correctly; read as revenue or GPU-hours, they overstate both. Spending data shows the gap: in Ramp’s AI Index, token volume rose 37% between the weeks of September 25 and October 23 while token spending rose 9%.⁴² Against a chat-era task of 100,000 tokens, the agent task is five times the tokens, 1.7 times the dollars and 1.5 times the compute. Move a whole fleet onto work like this and the token chart rises 5.6×, revenue 1.2×.¹⁴

Coding soaked up the first wave of supply for three reasons. It is iterative: the model can try, run, and try again. It is single-player: an engineer works in an isolated worktree and merges when done. It is verifiable: compilers and tests catch errors. Anthropic says over 90% of the contributions to its own codebase in September 2026 were written by Claude.¹⁵

That demand is real, and it is maturing. Revenue grew 132% quarter on quarter in the first quarter of 2026 and about 145% in the second; the filing’s own comparison, third-quarter revenue at 86 times the fourth quarter of 2024, implies roughly $16.9 billion and growth of about 46%.¹⁶ Each model generation cuts customers’ cost per task: 25% with Fable 5.1, 40% with Opus 5.5, by the company’s own account “a combination of a lower price and improved efficiency”.¹⁷ Concentration is easing but still high: the top five customers were about 21% of third-quarter revenue, down from 36% two quarters earlier.¹⁸

Cyber was meant to be the second wave. Most enterprises respond to AI vulnerabilities by locking down permissions and turtling, not by burning tokens rewriting code.

If completed coding work flattens, my working assumption is that 95% of incremental demand must come from non-coding enterprise work.¹⁹ That work sits elsewhere on the map: most of it is multiplayer, much of it is open-ended, and all of it must be governed.

Accepted work is the product of six fractions: the underlying opportunities, the share that is technically eligible, the share that can be verified cheaply enough, the share whose data and governance are ready, the share actually deployed, and the share that pays. Paid tokens are accepted work times tokens per outcome times the provider’s capture.²⁰ Multiply six fractions under one and you get a small number.

accepted work = opportunities × eligible × verifiable × governed × deployed × economic
paid tokens = accepted work × tokens per outcome × provider capture

The Work/Demand Lab behind this essay runs 1,000 real occupational task statements through those gates. In its default “Slop Chasm” scenario, a year out, 100 tasks are out of scope (science and physical operations), 98 fail the acceptance threshold, 228 fail the variable economics once review and rework are counted, 145 cannot cover a department’s implementation bill, and 429 are active. Of 7.7 billion monthly task opportunities, 284 million are attempted and 253 million accepted: about 3%.²¹

Three gates do most of the damage. Compounding error: a ten-step process at 98% a step is 81.7% end to end. Where errors do not compound, AI works: in customer service the vendor Intercom reports its agent resolving about two-thirds of conversations.²² Arms and legs: complex sales, physical logistics and cross-department approvals need shared state, accountability and people in the room. And the total workflow cost gate, which gets its own machine.

The cost of an accepted outcome is model tokens plus tools and infrastructure plus human review plus forward-deployed implementation plus error liability. Suppose tokens are 10% of that and everything else is 90%.²³ Cut the token price 40% and tokens per task 20%. The new total is 90% + 10% × 0.6 × 0.8 = 94.8%. The enterprise saves 5.2%. To get 2.5 times as much work out of a 5.2% price drop, demand would need an elasticity of 17.

Ten percent is generous. In the Lab’s default profile for a billing reconciliation, the model’s tokens cost 4 cents; human review costs $14.40, coordination $2, rework $1.90, expected errors 95 cents, the platform fee 25 cents. Tokens are a fifth of one percent.²⁴

This is why the ROI stories that do get reported are bounded workflows inside the systems that already hold the data and the approvals: Oracle, SAP, ServiceNow, Salesforce. Those platforms route and distill each step onto the cheapest adequate model. DeepSeek’s V4.1-Flash, updated in October, lists at 30 cents per million tokens in and $1.20 out and is tuned for long-horizon coding harnesses including Claude Code and Codex; Haiku 5.5 lists at ten cents in.²⁵ The frontier tier is the last thing a governed workflow reaches for. MBI’s own 2030 model needs about $200 billion of revenue from work that only frontier tiers can do, and asks who will pay for it.²⁶

Take a normalized fleet of 100 units: 25 for training, 15 held in reserve, 60 available for inference, of which 30 are paid today. That is 50% paid utilization and $5 billion a month of gross revenue, against a compute bill the model sets at $10.1 billion a month for 2027.²⁷

Now let a year pass with the model’s defaults. Serving throughput doubles. Tokens per completed task fall 20%. The realized price per token falls 40%. Coding, 60% of today’s paid tokens, stays flat in completed work. Non-coding work grows by a factor g. Paid tokens become 0.8 × (0.6 + 0.4g).

If non-coding work doubles, paid tokens rise 12%, and paid utilization falls from 50% to 28%, because the same tokens need half the GPU-hours. Gross revenue falls a third, to $3.4 billion a month. To hold revenue flat, non-coding work must grow 3.7 times. To hold utilization at 50%, 4.75 times. To cover the $10.1 billion bill, about 9 times. To fill all 60 inference units, 11 times.²⁸ The Lab’s default path, with its 1,000 gated tasks, takes paid utilization from 50% to 20% in year one and 8% in year two.²⁹

Even a full fleet does not pay at today’s prices. Selling all 60 inference units would bring in about $9.2 billion a month after channel fees, against $11.1 billion of compute and other costs. Break-even needs 121% paid utilization: more than the fleet can sell.⁴³

Pythia Cap, a bull, frames the same question bottom-up: Anthropic’s non-coding users pay about $75 a month. Would every knowledge worker? “Probably not.”³⁰

At June 30, 2026, Anthropic had $265.6 billion of fixed and minimum purchase commitments and $321.4 billion of leases not yet commenced. Customers had committed $10.8 billion of remaining performance obligations.³¹ The two numbers have different horizons and definitions, so this is not a coverage ratio. It is a direction: fifty-four dollars of promised cost for every dollar of promised revenue.

The named contracts come in two kinds. Four are “regardless of actual usage” or non-cancelable: Broadcom lease rent of about $161.2 billion, Google at least $114.5 billion, Amazon at least $110 billion, Microsoft at least $44.9 billion. Together, $430.6 billion. Three are up-to, conditional or cancellable: SpaceX.AI up to $120.5 billion, about $3.1 billion a month once capacity is delivered in December 2026 and cancellable by either party on 90 days’ notice after January and March 2027; Nscale up to $44.6 billion, fees only as capacity is delivered and conditional on Nscale’s financing; Lambda about $41.3 billion, most of it terminable if a facility lease is not secured. Together, $206.4 billion.³²

Against that, demand arithmetic is sobering. A hundred million workers running twenty AI tasks a day, netting two cents a task, is $880 million a month: under 9% of a $10.1 billion bill.³³ The reported bill is already climbing: compute and infrastructure expense was $5.6 billion in the second quarter of 2026, up from $944 million five quarters earlier, and SpaceX.AI alone adds $3.1 billion a month from December.³⁴ MBI’s model has the company losing money in 2027 as contracted compute arrives ahead of the revenue it may support. Sridhar’s analogy is the right one: sign a ten-year take-or-pay contract, then sell cargoes at spot.³⁵

When paid utilization lags in 2027, the lab does not fail. It had $131.7 billion of cash and short-term investments at June 30, a $3.2 billion undrawn facility, and the IPO on top.³⁶ It has two levers instead.

First, cancel the soft contracts. The SpaceX.AI door opens in January and March 2027 on 90 days’ notice; the Nscale and Lambda commitments accrue only as capacity is delivered, and for Nscale only if it is financed.³⁷ Cancellation pushes the GPUs, and the debt behind them, onto the neoclouds and their lenders.

Second, pay the hard contracts in paper. The filing discloses note facilities of up to $82 billion, the largest a $42 billion Broadcom facility “solely to finance rent”, with notes that convert into shares at or below the trailing market price.³⁸ Unpaid rent becomes equity. The public float absorbs that dilution with a minority of the votes: a single founder-held Class F share carries 50.1% of total voting power, and the Long-Term Benefit Trust appoints a majority of the board. Five percent of the fully diluted shares already sit in a Public Benefit Fund, with up to 2% more at each 50% rise in the share price.³⁹ Damodaran values the equity at about $1.475 trillion against a rumored $1.5–2 trillion range and calls the voting structure a value discounter.⁴⁰

This is my reading of how 2027 plays out, not a disclosed plan. The mechanism, though, is in the filing.