Irina Peskisheva is Research Editor and Editorial Partnerships Lead at Inside Deep Tech. She covers emerging technologies, with a particular interest in how technology becomes a business.
Ilya Shimchik, team principal of Constructor Racing, on second place at Imola, on the 10 TB the team has recorded over three seasons, and on the January 2027 rule that puts a deadline on autonomous-system validation.
LPO strips the module DSP and keeps the pluggable cage: LPO MSA 100G-DR-LPO for 800G over 500 m SMF, OIF linear interfaces, Cisco and Marvell shipping signals, and when DSP optics or CPO still win.
OCS puts programmable photonic paths where electrical packet spines burn watts: MEMS, robotic, liquid-crystal, piezo, and SiP switches, Google's Jupiter and TPU deployments, and the honest limits versus CPO and GPU fabrics.
CFET stacks nFET and pFET into one footprint after GAAFET: Samsung's 42 nm gate-pitch 2026 VLSI Best Paper, imec's A7 / ~2033 commercial call, mCFET vs sCFET, and why volume chips still wait.
Five driverless cars raced at Imola on 5 September, the first multi-car autonomous race in Europe. Kinetiz won because the two cars ahead of it stopped. The three failures are what made the day worth the money, and they are a shopping list of questions for anyone buying autonomy.
AECs put DSP retimers in copper OSFP/QSFP-DD assemblies for AI scale-out: Credo ZeroFlap 800G SPAN to 7 m, Marvell Alaska A 1.6T beyond 3 m, and why DAC, ACC, AOC, and CPO are different BOM columns.
SOCAMM puts LPDDR5X on detachable compression modules for NVIDIA Vera: up to 1.2 TB/s and 1.5 TB, JEDEC JESD328 SOCAMM2 in June 2026, and why many racks ship 96 GB modules instead of 192 GB.
MRDIMM (Multiplexed Rank DIMM) doubles the bytes per cycle on server DDR5 by muxing two ranks through on-DIMM MRCD and MDB buffers. Gen1 ships near 8800 MT/s on Intel Xeon 6 versus RDIMM 6400, with JEDEC Gen2 raw cards targeting 12,800 MT/s.
Every team at Imola runs the same computer: one NVIDIA RTX 6000 Ada and one sixteen-core AMD EPYC. What separates them is the software, and software is what closed the gap to a human driver from three and a half minutes to six tenths of a second.
Everything you need to understand GDDR7 in 2026: JEDEC JESD239 PAM3 signaling, up to 192 GB/s per device, Micron and Samsung shipping speeds, how RTX 50 and Rubin CPX use it, and why board-level GDDR is not a substitute for HBM on an interposer.
Everything you need to understand PCIe 6.0 in 2026: 64 GT/s PAM4, 256 GB/s x16, FLIT mode, FEC/CRC, L0p power, how it differs from NVLink and NVLink-C2C, where Blackwell still uses Gen6 for NICs and host attach, and why PCIe is not the GPU-to-GPU fabric.
In 2019 the estimate was 20 million noisy qubits and eight hours. In 2025 the same author, assuming the same hardware, put it under a million qubits and under a week. Nothing was built in between.
Unitree listed on 19 August and its filing is the first audited view of humanoid robot economics. Research and education took 73,60% of humanoid revenue in the first nine months of 2025. Industrial application took 9,01%, and between half and seven tenths of that is corporate guided tours.
The five most active investors in each of the four frontier markets, ranked on published fund size. Quantum has three dedicated funds in its top five. Robotics has none.
What quantum computing has actually demonstrated by 2026, what is only on a roadmap, and how to tell the two apart. Every figure traced to its source.
Twenty deep tech investors, ranked on the one number each firm publishes about itself: the size of its last announced fund, with the date and the backers behind it.
Ten inference accelerators compared on measured benchmark results where those exist, and on published specifications where they do not. Which is which is the finding.
IBM says it has more quantum computers deployed than the rest of the industry combined. Most of them sit in its own data centres. Seven of these twelve companies have machines standing on floors they do not own. Five have none.
