The slow clock.

AI is cutting the time it takes to design a chip. The time it takes to make one is set by calendars that AI does not run. Semiconductor money in 2026 has bought design and bought demand, and very little of it is aimed at the wait between them.

A silicon die held between finger and thumb, its circuitry catching the light

In July Synopsys said its new verification agent reaches validated RTL up to 50 times faster than traditional verification workflows. The product is still with customers for evaluation. Earlier in the year Nvidia’s chief scientist said AI had cut a ten-month, eight-engineer job, porting a standard-cell library, to a single night. These are the vendors’ own figures, and no common benchmark yet tests them.

Now book the part. Europractice, a leading shared-run route to prototype wafers for smaller companies, offers imec’s active silicon photonics process twice in 2026, with tape-outs on 22 April and 21 October. A design that misses the April deadline by a week waits six months for the next start. Masks, fabrication, packaging and test come after that.

Two clocks

A hardware company runs on two clocks. The design clock is set by people and compute, and AI is speeding it up. The making clock is set by shuttle calendars, mask shops, packaging queues and customer qualification. Those are capacity decisions taken by other organisations, and faster software does not move them.

As the first clock speeds up, the second becomes most of the schedule. A programme moves at the speed of its slowest loop, and the slowest loop is now the physical one.

Two clocksTurn up the AI. One clock answers.

The design clock

People and compute. AI is speeding it up.

The making clock

Shuttle calendars, mask shops, packaging queues, customer qualification.

NoneHeavy

The programme: design plus making

DesignMakingThe making is 87% of the schedule

A programme moves at the speed of its slowest loop.

Illustrative proportions. The making loop is held where it is; the design loop shrinks as AI is applied to it. The dashed frame is the programme before any AI.

Runway

Investors count runway in months. A hardware founder spends it in loops: design, make, test, learn, redesign.

Eighteen months of runway is about 78 weeks. On a process offered twice a year, that is three tape-out slots at most, and every slip costs six months. A quarterly run gives six slots. A monthly one gives eighteen.

Runway loopsEighteen months of runway, counted in tape-out slots

Twice a year3 slots

Quarterly6 slots

Monthly18 slots

Week 0Week 78 · 18 months of runway ends
  • Tape-out slot · miss one and the design waits for the next
  • The wait after a one-week slip

Slots spaced evenly on each calendar, with the first in week 8. A design ready one week after that slot waits for the next one on its lane.

Runway is counted in loops. Doubling the number of runs a year gives a founder the same number of attempts as doubling the round, with no dilution to anyone.

What the money bought

It has been a strong year for UK chip finance. In June the government set out a £1.1bn AI hardware plan, including a £150m advance commitment to buy novel inference chips from British firms once they meet agreed benchmarks. The British Business Bank made a commitment of up to £150m, the largest in its history, to a new UK AI hardware fund led by Playground Global. In July it also made up to £500m of existing guarantee capacity available, for twelve months, for lending against intellectual property.

The pace has not slowed this month. On 11 November Fractile announced a $720m Series C for its inference chips, with the Sovereign AI fund among the new investors. A day earlier the AI minister announced a Sovereign AI investment in Tendrils Compute, as part of a round to reinvent the computer processor. Both are good news for British design. Shipping is where the slow clock sets the pace.

Sort those by clock. Equity pays for design teams. The advance commitment pays for demand. The IP guarantee lends against the patent, which describes the design. Each is useful. Each pays for the fast clock or for the finish line, and leaves the slow clock where it was.

Sort those by clockWhat the 2026 money pays for

The fast clockDesign

EquityPays for design teams
IP guarantee, up to £500mLends against the patent, which describes the design

The slow clockMaking

Left where it was

The finish lineDemand

Advance commitment, £150mPays for demand

DSIT’s 2026 sector study counts about 700 UK semiconductor companies, 295 of them dedicated to the sector, and finds that a small number of large firms account for most of the revenue. The rest are mainly small, and small companies prototype on shared runs, on somebody else’s calendar.

The floor

At techUK’s session on financing the future of semiconductors on 10 November, the panels answered the money questions well. The questions about making came from the audience: a graphene company that could not raise money to scale, a question on how Arm moved from selling designs towards selling products, and a question on what government could do beyond finance. An opening slide had put manufacturing on a level with funding. No panellist was introduced as speaking for manufacturing.

The second panel was titled “Beyond the cheque: partnerships, commercialisation, exports and R&D”. It was the right question, and the list has room for one more. The item to add beyond the cheque is Manufacturing Engineering.

The track record

The keynote from the UK Semiconductor Centre set out the previous generation of British chip companies.

The track recordEleven British chip companies, and where they went
Eleven British chip companies by founding year. Ten lines leave for the buyer’s country: seven to the United States, two to China, one to Japan. EnSilica stays in the UK, independent. UKOverseas 1984WolfsonCirrus Logic · 2014 1985ImaginationCanyon Bridge · 2017 1990ArmSoftBank · 2016 1997ARC InternationalVirage Logic · 2009 1998CSRQualcomm · 2015 1999Element 14Broadcom · 2000 2000CamSemiPower Integrations · 2015 2000picoChipMindspeed · 2012 2001EnSilicaStill independent · AIM 2022 2002IceraNvidia · 2011 2010NeulHuawei · 2014 Japan1 company United States7 companies China2 companies

To the United States 7

  • Wolfson, founded 1984: Cirrus Logic (US), 2014
  • ARC International, founded 1997: Virage Logic (US), 2009
  • CSR, founded 1998: Qualcomm (US), 2015
  • Element 14, founded 1999: Broadcom (US), 2000
  • CamSemi, founded 2000: Power Integrations (US), 2015
  • picoChip, founded 2000: Mindspeed (US), 2012
  • Icera, founded 2002: Nvidia (US), 2011

To China 2

  • Imagination, founded 1985: Canyon Bridge (China-backed), 2017
  • Neul, founded 2010: Huawei (China), 2014

To Japan 1

  • Arm, founded 1990: SoftBank (Japan), 2016

Stayed in the UK 1

  • EnSilica, founded 2001: still independent; listed on AIM, 2022
CompanyFoundedWhat it didWhat happened next
CSR1998Bluetooth connectivity chipsLSE-listed (2004); sold to Qualcomm (US), 2015
Wolfson1984Audio and mixed-signal chipsLSE-listed; sold to Cirrus Logic (US), 2014
Imagination1985PowerVR graphics processor IPLSE-listed; sold to Canyon Bridge (China-backed), 2017
Arm1990Processor architecture IPLSE-listed (1998); sold to SoftBank (Japan), 2016
ARC International1997Configurable processor IPLSE-listed; sold to Virage Logic (US), 2009
Element 141999DSP and broadband chip designPrivate, VC-backed; sold to Broadcom (US), 2000
Icera2002Software-defined mobile modemsPrivate, VC-backed; sold to Nvidia (US), 2011
picoChip2000Small-cell processorsPrivate, VC-backed; sold to Mindspeed (US), 2012
CamSemi2000Power management chipsPrivate, VC-backed; sold to Power Integrations (US), 2015
Neul2010IoT connectivity chipsPrivate, VC-backed; sold to Huawei (China), 2014
EnSilica2001Mixed-signal ASIC designStill independent; listed on AIM, 2022

Since 2023Arm listed on Nasdaq alone (2023), private Graphcore sold to SoftBank (2024) and London-listed Alphawave sold to Qualcomm (2025).

Source: UK Semiconductor Centre, “The track record”, Financing the future of semiconductors, techUK, 10 November 2026, slide 7. Illustrative examples; the slide cites company histories and acquisition records.

Ten of the eleven were sold overseas. The slide drew two priorities from that: growth capital for private firms and long-term backing for listed ones. Both are about money. Every company in the table was known for its designs or its IP, and a buyer acquires a design business by acquiring its files and its team. A company whose making is written down and running on a UK line gives a buyer more to move, and gives a British owner more reason to stay.

Shortening the loop

Three changes would put the slow clock on the table, and each builds on an instrument that already exists.

Publish the clock. Any fab, packaging house or open-access line that takes public money should publish its turnaround from frozen design to tested part, quarter by quarter. Shuttle calendars are public. Packaging and test lead times mostly are not, and founders and investors cannot price a wait they cannot see.

Tie tranches to the making. Infrastructure lenders release funds against milestones certified by an independent engineer. Venture-backed hard-tech companies seldom have that role filled. Tranches tied to signed manufacturing gates would let a funder reward a shorter loop directly.

Buy the slots as well as the chips. Reserved runs for UK companies on lines inside existing UK fabs, committed in advance, would take the shared calendar off the critical path. The £150m commitment shows government will promise to buy a product before it has been proven. The same reasoning applies to the slots that make it.

Beyond the cheque, Manufacturing Engineers take weeks out of the loop. They qualify a second packaging route before the first one slips. They run the second lot and explain why it differs from the first. They write the route down so that another line can run it.

Kaipability works at this interface. For a funder, it provides the independent view that a manufacturing gate has been met, and it does not do that for a company it is also helping to close the gap. Funds, founders and the authors of techUK’s report to government who want the slow clock counted should talk to us.

Sources

Q&A

Questions this dispatch answers.

Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.

What is the slow clock in chip development?
The slow clock is the making loop of a hardware programme: shuttle calendars, mask shops, packaging queues and customer qualification. They are capacity decisions taken by other organisations, so AI that speeds up chip design does not move them, and as design gets faster the making becomes most of the schedule.
Why does AI not shorten the time it takes to make a chip?
Because AI speeds up design, which runs on people and compute, while making runs on shared calendars. Europractice offers imec’s active silicon photonics process twice in 2026, so a design that misses a tape-out by a week waits six months for the next start. A programme moves at the speed of its slowest loop.
If AI lets anyone design a chip, where does the advantage go?
To whoever controls the making. As AI spreads, chip designs will come from everywhere, and every one of them queues for the same shuttle runs, mask shops and packaging lines. The advantage goes to the lines that publish and shorten the wait from frozen design to tested part, and to funders who tie tranches to signed manufacturing gates and reserve runs in advance.