A semiconductor company does not sell a chip. It sells a line item in a bill of materials for a product that does not exist yet, to a customer whose own launch date decides when the money arrives. Everything strange about semiconductor go-to-market follows from that one fact, and almost every instrument a first-time commercial leader imports from software assumes the opposite. The practice behind this note lives at GTM consulting for deep tech and hardware.
The money arrived. The commercial architecture did not.
Crunchbase counted roughly $10.7 billion into seed-through-pre-IPO semiconductor rounds by June 10, 2026, on pace to pass 2025. MatX raised a $500 million Series B in February led by Jane Street and Situational Awareness. Ayar Labs raised $500 million in a March Series E led by Neuberger Berman, with AMD Ventures and Nvidia on the cap table as strategics. Cerebras went public and raised over $5 billion. The sector is not short of capital, and it is not short of silicon. It is short of commercial functions built for the way silicon is actually bought.
That gap is where the generic go-to-market consultant does the most damage. A playbook that works on a software purchase assumes the buyer can decide, buy and deploy inside one budget cycle, and that the seller controls the close. On a socket sale, none of the three holds. The general version of that argument is why hardware doesn't sell like SaaS. This note is the semiconductor case, which is the most extreme case in hardware.
A design win is not a booking. It is an option on someone else's product reaching volume.
Four instruments that break on a socket sale.
One: pipeline coverage. Coverage is arithmetic, not folklore. The required qualified pipeline works out to the inverse of the win rate on that pipeline, and at manufacturing's 19 percent median win rate that implies roughly 5.3x (Hardware GTM Benchmarks 2026, derived). The arithmetic is fine. The inputs are not, because in a semiconductor company the numerator and the denominator are usually measured in different years. Win rate gets measured on design-win attempts closing this year. The quota it is covering is revenue that will be recognised when the customer ramps. Any coverage multiple built by dividing one by the other is a number with no referent.
Two: the comp plan. The manufacturing baseline deal runs $48K on a 124-day cycle at a 19 percent win rate (Digital Bloom), and serious capital-equipment and committee-gated purchases run 9 to 18 months. Semiconductor design cycles sit past even that. Industry write-ups on automotive silicon commonly describe production revenue trailing the design win by roughly 12 to 24 months on qualification cycles alone. That is a market observation rather than a SignalForge benchmark, and the exact figure is less important than the shape: a commission plan that pays on recognised revenue pays a seller for the customer's product schedule. It is the single fastest way to lose the person who won the socket, because they leave before the money they earned arrives.
Three: the forecast. Converting design wins to revenue on a fixed attach assumption produces a forecast of the customer's execution, presented as a forecast of yours. Sockets get cancelled. Programs slip. End products fail in their own market and take the silicon with them. The number is not wrong because the assumption is too aggressive; it is wrong because it is unexamined and carried forward year over year without ever being scored against what actually shipped.
Four: single-threading. The modal B2B buying committee runs 6.3 to 6.8 people (Gartner and 6sense composite), and a semiconductor design-in assembles a particularly awkward version of it. The hardware engineer who wants your part. The firmware lead who has to live with your software development kit. Component engineering, who decides whether a single-source part is even allowed. Quality, who owns qualification. Procurement, who negotiates against a second source you may not have. And in most cases the customer's own customer, whose requirements set the spec everyone else is arguing inside. Selling to the engineer alone under-covers that room by five seats. The anatomy is in the hardware buying committee.
The evaluation board is a pilot, and it converts like one.
Median pilot-to-production conversion across hardware is 12 percent (IDC 2025). Chip companies rarely call their eval boards pilots, so they rarely apply the discipline, and the conversion behaves the same way regardless of the vocabulary. An eval board shipped to an engineer who asked nicely is the semiconductor version of the pilot that dies at the budget handoff: someone technical is satisfied, and nobody with a program budget was ever in the conversation.
What converts an eval is what converts any pilot, and it is decided before the board ships rather than after it works. A named program the part is being evaluated for, with a start-of-production date attached. A named person who owns the decision to design it in, distinct from the engineer running the bench. The second-source position stated plainly, because component engineering will raise it and the honest answer is better than a surprised one. The qualification burden agreed in writing, including who pays for it and what the schedule is. Non-recurring engineering terms that do not quietly obligate a roadmap. The design rules are in pilot to production; the semiconductor translation is only the vocabulary.
The pattern the good companies run is unglamorous: fewer evals, each one attached to a program. A high eval-board count is a vanity metric with shipping costs. It measures curiosity, and curiosity is free.
What to build instead: four moves.
If you run a semiconductor company between $1M and $20M in revenue, the fix is not a bigger sales team ahead of a documented motion. It is an instrument set that matches the transaction. Four moves, in order.
Instrument the design win as a stage, not an anecdote. A design win is only real when three fields are filled: the socket, the customer program it sits in, and the expected start of production. A design win with a blank program field is a relationship, and a design win with a blank production date is a hope. Companies that will not enforce those three fields cannot forecast, and no consultant can fix a forecast built on records that do not exist.
Run two pipelines and never merge them into one board slide. The design-win pipeline is what the commercial team controls: opportunities to be selected. The revenue pipeline is what the customer controls: won sockets moving to volume. They have different conversion rates, different cycle times and different owners. A board shown a single number gets a number that means two things, and the ambiguity always resolves in the optimistic direction. Two charts is not a reporting preference, it is the difference between a company that knows why it missed and one that does not.
Pay on the event the seller controls. Design-win credit at design-win time, with a revenue component that follows the ramp. The exact split is a company decision and depends on runway and on how much of the ramp risk sits inside the customer rather than inside the product. What is not a company decision is the principle: a plan that pays only on recognised revenue makes every seller a volunteer for one to two years, and the good ones will not volunteer twice.
Make the reference design citable, not gated. 51 percent of B2B buyers now start vendor research in an AI assistant rather than a search engine (G2, March 2026, survey of 1,076 buyers). For a chip company that is not a marketing observation, it is an availability problem. If the reference design, the evaluation results, the software development kit documentation and the integration path are behind a form, they are absent from the first screen of the modern design-in process, where the engineer is asking an assistant which part solves their problem before anyone in your company knows the program exists. The method that turns technical evidence into pipeline on purpose is Proof to Pipeline.
A funded sector is not a buying sector. The $10.7 billion is investor conviction about AI infrastructure, not evidence that design-in cycles have shortened or that the buyers on the other side of the table have more budget. Cerebras raised over $5 billion in its IPO and its shares were down about a third from the first-day close five weeks later (Crunchbase, June 10, 2026). Capital markets and customer markets are different markets and they turn at different times.
The structural half of this note does not move with the cycle. Silicon has been bought through a design win, a qualification and a customer ramp for decades, and it will still be bought that way when the current wave recedes. An instrument set that matches the transaction wins in both markets.
Three questions, answered straight.
What does semiconductor GTM consulting actually cover?
It covers the commercial architecture around a socket sale rather than generic sales training. The work is defining the design win as an instrumented stage with a named socket, program and start-of-production date; separating the design-win pipeline from the revenue pipeline so a board is never shown one number that means two things; building a comp plan that pays on the event the seller controls; and making the reference design, evaluation data and integration path citable, because 51 percent of buyers now start vendor research in an AI assistant (G2, March 2026). The firm-level version of that engagement is described on the GTM consulting page; the earlier-stage version is go-to-market consulting for hardware startups.
Why do semiconductor sales forecasts miss even when design wins are up?
Because a design win is an option on someone else's product, not a booking. Revenue arrives when the customer reaches production volume, which is their schedule and not yours, and a share of won sockets never ship at all. A forecast built by converting design wins to revenue on a fixed assumption is a forecast of the customer's execution. The fix is two pipelines with two conversion rates measured separately, and an attach assumption that is written down and scored against outcomes rather than carried forward unexamined.
Is an evaluation board the same as a hardware pilot?
Structurally yes, and it carries the same conversion problem. Median pilot-to-production conversion across hardware is 12 percent (IDC 2025), and an eval board handed to an engineer with no production-intent decision attached is the semiconductor version of the pilot that dies at the budget handoff. What converts an eval is the same thing that converts a pilot: a named program it is being evaluated for, a named person who owns the production decision, and the second-source and qualification terms agreed before the board ships rather than after it succeeds.
Keep reading.
This note sits inside a published method. Start with the page that ties it together, or go straight to the piece that names your stall point.