The throughline in this issue is the conversion step — the distance between a capability that works and one that ships. Entanglement survives 21 kilometres of open air, the US starts rebuilding an enrichment base it let fall from 64% of world capacity to under 1%, and six separate robotics papers land on the same wall between a trained model and a working robot. None of it is new science. All of it decides whether the science becomes a product.

BCI.INTEL · Brain-Computer Interfaces

RMB 6.7 billion across 72 deals in six months — 2.5 times the prior comparable period — is China's brain-computer interface funding run rate. The figure framed an international BCI conference in Shenzhen organised by Hong Kong University and the Advanced Biomedical Instrumental Center, where the argument from investors was that Europe's scientific base is not the constraint; venture infrastructure and conversion are. One participant put China's 2030 domestic market share at 40% or more, roughly double conservative estimates — a motivated number from an invested party, not independent analysis. Treat the share claim carefully. The deal count is the harder signal.

Read the full analysis on BCI.INTEL →

HUMANOID.INTEL · Humanoid Robotics

Six separate papers converged on one failure mode: a vision-language-action model cannot learn a new skill without degrading the skills it already had. OrthoSkillVLA applies gradient-informed subspace constraints and a lightweight mixture-of-experts decoder; a self-demonstration fine-tuning method preserves zero-shot priors while training on new hardware; Remember Smarter lifts pi0 on LIBERO-Plus from 53.6% to 70.6% without expanding context; CoRe recovers frozen policies at inference with no retraining and no failure data. A survey names the adjacent problem — the embodiment gap, the cost of moving a trained model onto a specific body — and TGL calls the whole thing the retraining tax. Twelve of seventeen humanoid papers were VLA work; half went at this one wall. A foundation model that must be retrained per task and per robot is a research artefact, not a product.

Read the full analysis on HUMANOID.INTEL →

QUANTUM.INTEL · Quantum Computing

21 kilometres of turbulent ground-layer atmosphere now carries entanglement-verified photon pairs — the first permanent free-space optical quantum link on US soil. Brookhaven National Laboratory and Stony Brook University transmitted single photons and entangled pairs across 13 miles of low-altitude air, the hardest optical environment short of crossing a mountain range. Because the no-cloning theorem forbids amplifying an unknown quantum state, the link has to hold phase coherence and entanglement fidelity through that turbulence rather than repeat the signal. Fibre is the solved case. Free space is what eventually connects buildings, and satellites, without trenching a route between them.

Read the full analysis on QUANTUM.INTEL →

SMR.INTEL · Small Modular Reactors

The US held 64% of global uranium enrichment capacity in 1985 and under 1% by 2024. Russia holds 43%. Four enrichers laid out rebuild plans at the American Nuclear Society's Global 2026: Urenco is installing 2.1 million SWU of new gas centrifuge capacity at Eunice, New Mexico, taking the site past 7 million SWU by 2036, a site the Energy Secretary attended in person to break ground; Centrus holds a $900 million DOE agreement to expand HALEU output at Piketon; Global Laser Enrichment is targeting construction in 2027; General Matter has leased 100 acres at the former Paducah plant. Urenco's own government-affairs head named the catch — advanced reactor fuel demand is still not firmly contracted. This capacity is being built ahead of committed offtake.

Read the full analysis on SMR.INTEL →

SYNBIO.INTEL · Synthetic Biology

A boy died in Huidagene's Duchenne muscular dystrophy gene-editing trial, and the field is auditing the trial design in public. Independent gene-editing and muscle-disorder specialists reviewed the presentation data for STAT and split hard — from "not sure it was unreasonable to try" to a characterisation of the underlying data as garbage. The company's former chief executive had presented the early trial results at the American Society of Gene and Cell Therapy meeting three months before the death. The scrutiny is landing on study design, dosing rationale and data quality rather than on the editing chemistry, which is the more consequential place for it to land.

Read the full analysis on SYNBIO.INTEL →

ORBITAL.INTEL · Space Economy

The National Reconnaissance Office ended a four-year evaluation and moved HawkEye 360 into operational radio-frequency intelligence. The award under the Commercial Radio Frequency Capabilities Augmentation programme, announced August 17, closes an assessment that began in September 2022 with study contracts to six companies; the contract value was not disclosed. HawkEye 360 flies more than 30 satellites in three-satellite clusters that geolocate emissions from radars, ship transponders and communications gear, and booked $49.8 million in Q2 revenue with a record $21 million international, up 134% year on year. Commercial RF intelligence just became a procurement line item rather than an experiment.

Read the full analysis on ORBITAL.INTEL →

Also moving: Epicrispr closed an oversubscribed $90M Series C to take EPI-321 into a pivotal FSHD trial · Doosan Enerbility took a second Laiyang forging order — eight components up to 130 tonnes each · Elve qualified a 100W millimetre-wave TWTA family for spaceflight · Brainsmart is seeking RMB 200M on a 256-channel ECoG array that cleared NMPA type-testing

That's the issue. If a sector moved and we missed it, reply and tell us — a real person reads every response.

Know someone who follows one of these six fields and ignores the other five? Forward this to them.

The Deep Tech Dispatch is published by DeepTech.Intel. Subscribe · deeptechintel.ai