Home›What moved the last eight weeks

What moved

The changes in the world view over the last eight weeks: which beliefs rose or fell, and what changed in each market's record.

8 Nov1 Nov25 Oct18 Oct11 Oct4 Oct27 Sep20 Sep13 Seprosefell 8 Nov25 Oct11 Oct27 Sep13 Seprosefell
Fig. 1The last eight weeks, left to right. A belief that rose stands above the line and one that fell is dug below it, as tall as the points it moved; the narrow blocks are the changes to markets' records each day, as tall as there were changes. Washed: No single photonics material wins, the largest move.
Beliefsmoved at the last recompute
Marketswhat changed in each record
The other 44
  • 23 Oct 2026Confidential computing. Two hosting platforms lost their customers secrets from the inside this week. Confidential computing is built for exactly that case, and neither was using it.
  • 23 Oct 2026Cyber insurance repricing. Cyber insurers are still cutting prices while their losses go up. That gap closes one way, and it ends with the smallest firms uninsured.
  • 23 Oct 2026Decode chips. Light cannot fix decode's weight-fetch problem, but a chip that keeps its weights in memory bonded under the optics has already fixed it, so the photonic question becomes the energy cost of turning weights into light.
  • 23 Oct 2026Hardware attestation. I score hardware attestation at 20. Europe's rules are being written around outcomes, not chips, and Intel has made itself the neutral referee the startups were hoping to be.
  • 23 Oct 2026Hardware attestation. Three poisoned updates in a week, all of which signed build provenance would have flagged. That fix is software and free, which is why the hardware end stays with the chip vendors.
  • 23 Oct 2026Hardware memory safety. I score hardware memory safety at 20. It fixes the bugs AI finds fastest, but almost none of this month's breaches came through that door, and nobody is being made to buy the chips.
  • 23 Oct 2026Hardware memory safety. Memory-safe code would have stopped the PHP bug and maybe two others out of 24. The rest came through stolen passwords and poisoned updates. Memory safety is one fix for one class, not the fix.
  • 23 Oct 2026Non-human identity and secrets. The biggest cause of last month's breaches was keys that software holds, not people's passwords. It is a real and growing market, and the incumbents bought the startups before the rule that would force the spend has even arrived.
  • 23 Oct 2026Photonic test and metrology. Testing photonic chips is a real bottleneck for co-packaged optics, and the test-equipment incumbents have already bought or built their way into it. The new companies in the category have raised almost nothing.
  • 23 Oct 2026Software supply-chain integrity. Three poisoned software updates in one month, all of which signed build provenance would have flagged. The signing is free; the money is in components someone rebuilds clean every day.
  • 22 Oct 2026Agent identity. Google now makes apps declare whether they are AI agents before they can reach Gmail or Drive through its new agent interface. The platforms are building agent identity into their own logins.
  • 22 Oct 2026Decode chips. The biggest independent HBM-free inference-chip vendor chose a $1.5B round over a sale to Intel or Qualcomm.
  • 22 Oct 2026Decode chips. A $6.2B valuation on an inference-chip company with no disclosed customer is a bet on the category, not a measurement of it.
  • 22 Oct 2026Fog computing middleware. Nvidia's own filings flag that open models could end up on rival hardware; software that makes chips interchangeable is how that happens.
  • 22 Oct 2026Fog computing middleware. By buying Hugging Face, Nvidia is moving to own the front door to open models, which could make its hardware the default even without requiring it.
  • 22 Oct 2026Functional safety silicon for robotics. A Mercedes-backed safety-chiplet start-up could not raise a productisation round within a year of spinning out.
  • 22 Oct 2026Hyperscaler ASICs. Nvidia is funding the buyers who will never build their own chip, AI clouds and smaller labs, because its biggest customers are all moving onto their own silicon.
  • 22 Oct 2026Hyperscaler ASICs. One broker says Google''s next TPU is three weeks late on packaging yield at Intel, which would make packaging the gate at a second foundry, not just TSMC.
  • 22 Oct 2026Hyperscaler ASICs. If Intel''s packaging yield on a flagship external customer is 45-50% in late 2026, the second source of advanced packaging arrives later than the plans assume.
  • 22 Oct 2026Hyperscaler ASICs. Google''s warrant to Marvell pays out per $500M of purchases and commits to none of them.
  • 21 Oct 2026Industrial wireless IoT. The big metering numbers attached to the NR+ ecosystem belong to an older proprietary mesh, not the new standard; NR+ certification opens mid-2027 and one chip vendor still supplies the market.
  • 16 Oct 2026Discovery metrology. Faster electron microscopy through sparse sampling is reaching market as software the microscope makers embed, which means the speed-up mostly accrues to the instrument incumbents unless the software company keeps a per-unit royalty.
  • 16 Oct 2026Optical interconnect. The man who sells the switches says copper stays inside the rack and the pluggable keeps the volume until at least 2028. Co-packaged optics is a keynote, not yet a product line.
  • 16 Oct 2026Photonic light sources. Lumentum's CEO says he ships a third less than customers want and is sold out through 2027; a forecaster says the shortage clears this year. One of them is wrong by December.
  • 15 Oct 2026Agentic compute silicon. The memory tier for AI agents is now named on an earnings call, and the companies naming it are selling commodity DRAM and SSDs into it. The demand is real; the question is whether anything but software ever gets to sell into it.
  • 15 Oct 2026Industrial wireless IoT. A licence-free industrial radio standard has its first products in the field, but no shipment data yet, a single silicon supplier, and a certification programme that now looks like mid-2027.
  • 15 Oct 2026Optical circuit switching. Two European optical-switching startups each raised $125 million in six weeks. In twenty years no company selling only an optical switch has become a large independent business, and investors are now betting heavily that this time is different. The money proves appetite, not outcomes: one of the two names no customer, and neither has said how it handles the timing re-lock that has held the category back.
  • 15 Oct 2026Optical interconnect. The best-funded new optical-interconnect company of the year raised over $700 million, and the first thing it shipped to a hyperscaler was a plug-in transceiver. Co-packaged optics is on its roadmap, after an intermediate step. That is the order I expected: the plug-in module keeps selling while the more integrated options arrive in stages.
  • 15 Oct 2026Optical interconnect. Nvidia, the company most associated with putting optics right next to the switch chip, has just co-designed a new plug-in optical module with Cisco for the next speed generation. The plug-in transceiver is not dying on schedule, because the customers who have to swap failed parts in the field still want it.
  • 15 Oct 2026Optical interconnect. Nvidia's co-packaged optical switch is in production with a named five-company supply chain and six cloud customers lined up. I had bet it would slip well past its date, and that bet now looks likely to lose. The broader call, that optics-next-to-the-chip stays inside Nvidia's and Broadcom's own systems for a few more years, is untouched.
  • 13 Oct 2026In-package cooling. Counted once, the market for cooling inside an AI-chip package is a few hundred million dollars in 2030, not the billion-plus you get by adding up the converter, in-die and interface stories, because they compete for the same heat and the package owners keep much of it.
  • 13 Oct 2026Space-based datacentres. I rebuilt the orbital-datacentre model from the ground up and checked it against SemiAnalysis, Mach33, Starcloud and Google. Tokens from orbit cost about 8x Earth today on hardware that has actually flown, about 2x even if launch were free, and about 2x in 2035 if every paper-design radiator and solar array matures on schedule. Tokens per watt is a chip property, not a location property. Two things everyone gets wrong: dawn-dusk orbits still have eclipses, and GPUs cannot run hot enough to shrink the radiators the bull case needs.
  • 12 Oct 2026Light-controlled therapeutics. Light-controlled therapy is two or three years from its first US approvals outside the skin, and all of them are in places light already reaches: the eye, the bladder through a scope, a tumour lit by a laser.
  • 12 Oct 2026Light-controlled therapeutics. The money in light-controlled therapy is betting on geographic atrophy and an IPO window after the first optogenetic approval in 2027, not on light as a category.
  • 12 Oct 2026Nanotube electronics. Nanotube electronics has one real revenue line today, a fab consumable (EUV pellicles), and it shrank this year; the transistor story is now mostly a Chinese state programme.
  • 12 Oct 2026Photopharmacology in scope-reachable tissue. Light-switchable drugs skip gene therapy, but they still need light to get in. The useful question is which tissue a scope or catheter can light, not which molecule is cleverest.
  • 12 Oct 2026Quantum processor hardware. The US government has narrowed its fault-tolerant quantum computer bet to six builders. That makes the machine market bigger and the list of plausible winners shorter at the same time.
  • 7 Oct 2026Biocomputing. Biocomputing is now a Y Combinator category, with two companies in one batch. The money is moving faster than the science: still no customer paying for compute on living neurons.
  • 7 Oct 2026Neuromodulation. The expensive part of neuromodulation is the implant. Focused ultrasound can now reach deep brain targets without surgery, and an at-home brain stimulator got FDA approval on a single remote trial. If ultrasound wins one proper randomised trial, the economics of the whole category change.
  • 7 Oct 2026Space power beaming. Power beaming between spacecraft has its first in-orbit test this month, but the market for buying power in orbit still looks like a few hundred million dollars at the end of the decade.
  • 6 Oct 2026Low-power edge compute. A former head of a major earbud and wearables chip business says there is no burning need for new edge AI chips: today's earbud chips already run always-on voice and noise cancelling on a few milliamps, so most edge-AI startups are solving a problem the incumbents have already solved. The open questions are software (when a camera wearable should capture and send) and robotics.
  • 6 Oct 2026Robot autonomy. One former chip executive's view: NVIDIA has historically been weak at the edge, and robotics is the one area that might force a genuinely new edge chip architecture. It is a contrarian read against NVIDIA's current push into robot compute, and still unproven.
  • 1 Oct 2026Emerging non-volatile memory. The oldest ReRAM startup now builds on TSMC's ReRAM macro instead of its own device. In embedded memory, the foundry's process tends to beat the startup's.
  • 22 Sep 2026Agent identity. Amazon blocking Meta's shopping agent shows the big merchants, not standards bodies, are deciding which AI agents get to buy.
The referencemost recently revised

Members get every score, call and timing change, with the reasoning, in the changes feed and a weekly email.