01 / Build This Weekend
Inhaled mRNA vaccines could finally give you mucosal immunity without a needle, and without a nebulizer.
What just became possible
A November 2023 ChemRxiv paper demonstrated stable inhaled mRNA delivery via charge-assisted lipid nanoparticles, beating the nebulization-induced aggregation that killed every prior attempt. Spray-drying and thin-film freeze-drying produce inhalable dry powders with room-temperature stability. Microfluidic aerosolization platforms preserve LNP morphology end-to-end.
Why now
Standard injectable mRNA does not induce significant mucosal IgA in seronegative populations. Inhaled does. Public health agencies still buy billions of injectable doses every cycle because the alternatives have not been ready. The 2023 capability and the 2024-2026 LNP refinements finally make a needle-free, mucosal product viable.
What you'd build
A ready-to-use, nebulizer-free inhaled mRNA vaccine platform for respiratory pathogens. Buyer is CDC, ECDC, and sovereign-wealth-backed pandemic preparedness stockpiles. The mRNA vaccines market goes from $6.85B in 2026 to $25.64B by 2034 at 17.9% CAGR. The nasal/inhaled subsegment is the fastest growing slice.
Who's already moving
Ethris (SNIM RNA platform, Phase 3, EUA in Indonesia 2022). Arcturus Therapeutics (LUNAR platform, Kostaive EMA-approved, ARCT-032 in Phase 2). Orlance (MACH-1, NIH SBIR $3.3M for inhaled flu). Moderna and BioNTech are investing in mucosal platforms. CanSino has an inhalable adenovirus-based COVID vaccine in China. Bharat Biotech's iNCOVACC is approved in India.
The gap
Consistent lung deposition. Avoiding enzymatic degradation. Scalable spray-drying. The regulatory pathway for mucosal IgA endpoints is still evolving. Sanofi committed €400M/year and 450 hires to mRNA. AeroVax got $28M CAD from Canadian government. The funded race is on. The wedge is dry powder, hand-held inhaler, no cold chain.
02 / AI Makes This Possible
Neural networks can now be trained end-to-end to run directly on FPGA hardware with sub-20ns latency.
What just became possible
A February 2026 arXiv paper showed end-to-end differentiable training of neural networks that map directly to FPGA Lookup Tables. Standard model goes in, hardware-native binary circuit comes out. Sub-20ns deterministic inference latency on Xilinx Alveo or Intel Agilex. No expert tuning in the loop.
Why now
Vitis AI, FINN, and hls4ml exist but need expert tuning and do not guarantee sub-20ns. HFT desks spend millions on FPGA boards and C++/HLS specialists to get there manually. The 2026 paper closes the automation gap. Optiver grew headcount by 150 to 2,112 in 2024. Hudson River Trading went from 900 to 1,150. The talent is being paid $147k to $350k.
What you'd build
An automated FPGA synthesis platform: ONNX in, hardware bitstream out, sub-20ns SLA on a defined set of architectures. Sell directly to HFT and quant hedge funds as a development tool, or license to FPGA vendors. Ultra-low-latency trading infrastructure market: $84B in 2024 to $204B by 2030 at 14.4% CAGR.
Who's already moving
Magmio sells a full FPGA trading stack but needs manual customization. Lattice Semiconductor builds low-power edge FPGAs. RaiderChip develops hardware IP cores. AMD/Xilinx (Vitis AI, Alveo UL3524) and Intel/Altera (OpenVINO, Agilex) own the platforms but require expert tuning. Nobody offers automated, deterministic sub-20ns synthesis as a productized service.
The gap
Proving sub-20ns latency on arbitrary models, not curated benchmarks. Binary NN accuracy at the level HFT strategies need. FPGA resource constraints. DARPA ERI POSH and NSF CISE fund the research foundation. The talent moat is real: hire from Jane Street, Optiver, Citadel, or from AMD/Intel chip design teams.
03 / Deep Tech Bet
A new chip-fabric trick cuts memory traffic for AI models by 40% by treating communication like plumbing.
What just became possible
An April 2026 arXiv paper showed how to co-optimize dataflow with on-chip fabric collectives for Mixture-of-Experts models on tile-based accelerators. HBM traffic drops dramatically. Throughput goes up. For models like Switch Transformer with 3.2TB+ memory footprints, this is the difference between profitable and not.
Why now
MoE models now make up over 60% of open-source AI releases. Hyperscalers are projected to spend $450B on AI infrastructure in 2026 alone. The bottleneck shifted from compute to interconnect and memory. NVIDIA's CUDA dominance creates a high barrier, but the inefficiency of generic GPUs on sparse MoE inference creates a wedge.
What you'd build
A hardware-software co-optimized inference engine for tile-based AI accelerators serving MoE models. Drop-in compatibility with open-source MoE checkpoints. Sell to AWS, Azure, GCP, CoreWeave, Lambda. AI accelerator market: $33.69B in 2025 to $309B by 2030. AI inference market: $103B to $254B in the same period.
Who's already moving
d-Matrix raised $275M Series C in 2025 for 3D stacked digital-in-memory. Cerebras has wafer-scale for MoE. Groq has LPUs. NVIDIA H100/B200/Blackwell plus TensorRT-LLM and SGLang. AMD MI300X with MORI-EP for expert parallelism. Google TPU v5. Everyone is racing. The white space: drop-in MoE inference on tile-based accelerators with no model surgery.
The gap
12 to 24 month tape-out cycles and tens of millions in NRE. NVIDIA ecosystem lock-in. Drop-in compatibility for the LLM teams who do not want to rewrite serving code. NAIRR Pilot opens $100M+ in compute and platform access. DARPA put $18.6M into Princeton for AI-workload-optimized chips. The first vendor to deliver 2-5x cost/perf on real MoE wins one hyperscaler. That is enough.
A $19 test detects drug-resistant TB in 90 minutes. Most clinics still send sputum to a central lab.
What just became possible
The GeneXpert MTB/XDR cartridge at $19.80 detects isoniazid, fluoroquinolone, and second-line drug resistance in under 90 minutes. WHO prequalified. The November 2025 medRxiv evaluation in remote clinics showed it gives immediate treatment guidance without waiting for culture. The technology shipped. The deployment did not.
Why now
43% of MDR-TB cases were diagnosed in 2022. The other 57% are still treated empirically or not at all. The Global Fund put $1.4B into TB in 2021-2023 and $129M into public-private mix in 2024-2026. USAID/PEPFAR/UNITAID have a 10-year buy-down with Cepheid pricing tests at $9.98. The cartridges are subsidized. The on-site logistics are not.
What you'd build
A turnkey, solar-powered mobile diagnostic unit pre-loaded with GeneXpert MTB/XDR cartridges and telemedicine software. Pull up to a remote clinic, run the test, transmit results, ship cartridges by drone. Buyers are National TB Control Programs and large NGO implementing partners. TB diagnostics market: $2.94B in 2026 to $4.44B by 2035.
Who's already moving
Cepheid sells GeneXpert Edge (battery-powered) but not the integrated mobile unit. Molbio Diagnostics (Truenat, chip-based, portable) is strong in India and Africa. SD Biosensor (STANDARD M10 MDR-TB) launched 2024. Bioneer (IRON-qPCR) is in trials. The piece nobody owns: the integrated solar-mobile-telemedicine bundle as a single SKU.
The gap
Solar reliability in 40°C environments. Cartridge cold-chain and waste management. Integration with each country's HIS. Fragmented regulatory approval across high-burden countries. The Global Fund has emergency funding for mobile units. EDCTP funds field evaluations. The deal: bundle, sell to one country's NTP, prove operational economics, then replicate.
05 / Watch This Space
The average endometriosis diagnosis takes 2 to 3 years and costs $1,500 to $3,000 in laparoscopy. A blood test could replace both.
What just became possible
An August 2022 Research Square paper identified a multi-gene biomarker panel based on pyroptosis-related genes (IL-16, NLRP3, GSDME, CASP1/4/5/8) for non-invasive endometriosis diagnosis and severity prediction. Combined with machine learning over the panel, sensitivity and specificity start to rival surgery.
Why now
190 million women globally have endometriosis. 10% of reproductive-age women are diagnosed, on average years late. The NIH RADx Tech ACT ENDO Challenge offers up to $3M in cash prizes for non-invasive diagnostics. NIH estimates $30M annual research funding. The biomarker panels were ready in 2022. The clinical validation and FDA pathway are catching up now.
What you'd build
An FDA-cleared blood test kit for gynecologists. Sell to academic medical centers and high-volume fertility clinics. Endometriosis market is $1.7B in 2024 going to $5.6B by 2034 at 12.35% CAGR. Diagnostics is a meaningful subsegment, but the bigger win is unlocking earlier intervention before the disease compounds.
Who's already moving
Hera Biotech (acquired Scailyte for AI biomarker discovery, pre-FDA clearance). Aspira Women's Health (ENDOinform, AI-enabled blood test, pre-FDA). Kephera Diagnostics (CA-125 + BDNF panel, first non-invasive US launch). HerAnova (94.4% accuracy, FDA submission planned Q3 2026). The race is on but no one has cleared FDA yet.
The gap
Validation cohorts across ethnicities and disease stages. CPT codes and payer negotiation. Clinicians trained to trust a blood result over the laparoscopy they know. Hera, Aspira, Kephera, and HerAnova all have a path. Whoever pairs an FDA clearance with the first major payer contract owns the segment.
See you next week.
- Theis