01 / Build This Weekend
Runway just raised $315M at a $3B valuation. Pika and Luma are spending their rounds on H100 clusters. A 2-3x inference speedup is a startup-saving product.
What just became possible
A February 2025 arXiv paper introduced adaptive token-wise feature caching that accelerates diffusion transformer inference 2-3x with no model retraining. LeMiCa, SVG (Sparse VideoGen), and DiffSparse all demonstrated 2.33x to 2.5x lossless acceleration on CogVideoX and HunyuanVideo. Drop-in, no quality loss, no fine-tuning.
Why now
AI video software market: $1.81B in 2026 to $21.61B by 2034 at 46% CAGR. 23 new AI video startups received Series A or later funding in Q4 2025, the highest quarter on record. Series A/B teams over-provision 40-60% of GPU capacity and pay for inefficient baseline pipelines. AI engineer salaries are up 56% over non-AI roles.
What you'd build
A drop-in inference optimization library for video generation startups with 50+ concurrent requests and $50k+/month GPU bills. Hardware-aware integration with NVIDIA TensorRT for FP8. Pricing: percentage of monthly GPU savings. The buyer is the engineering lead at a $50M to $500M valuation video gen company who needs to extend runway by 6 months.
Who's already moving
NVIDIA TensorRT (Adobe Firefly production usage). LeMiCa (China Unicom, CogVideoX 2.5x). Sparse VideoGen (academic). DiffSparse (learned token sparsity). Most are research or model-specific. The whitespace: a productized library that wraps the academic methods and ships as a stable PyTorch wheel with enterprise support.
The gap
Architecture-genericity across DiT, Flow Matching, and the next two model families. Temporal consistency proofs. Compatibility across NVIDIA, AMD, AWS Inferentia. NSF SBIR up to $1.5M. AI Grant ($5k-$50k for open-source). Princeton AI Lab seed grants ($0-$60k). The first to land Runway, Pika, Luma as paying customers becomes the dependency.
02 / AI Makes This Possible
71% KV cache reduction with 2x throughput. Long-context LLM inference economics just collapsed.
What just became possible
A February 2026 arXiv paper validated residual-based KV cache compression delivering 71% memory reduction and 2x throughput. ChunkKV, AQUA-KV, and GEAR all maintain accuracy within 1% of baseline. Long-context inference cost (>32k tokens) drops 2x. Modern GPUs (H100, H200, AMD MI350X) support the sparse attention patterns natively.
Why now
Global AI infrastructure spend: $50B in 2024 growing 35% annually. Enterprise LLM market: $6.7B going to $71B by 2034 at 26% CAGR. 37% of enterprises spend $250k+/year on LLMs. Long-context inference is where the cost is rising 2x+ per query and KV cache eats 70% of GPU memory. Senior LLM engineers earn $195k to $350k.
What you'd build
A drop-in inference optimization library that auto-applies residual KV compression. Sell to LLM inference infrastructure providers (Together AI, Anyscale, Replicate) and enterprise AI platforms in legal tech and financial analysis. Pricing: per-throughput unit saved, with SLA on accuracy retention.
Who's already moving
NVIDIA TensorRT-LLM (de facto, includes FP8 and KV compression). vLLM (paged attention, default in production). Cerebras (launched Inference August 2024, 20x perf vs GPU, OpenAI 750MW deal). Neural Magic (acquired by Red Hat/IBM). LMCache (open-source, 15x throughput). No production-ready, auto-residual-compression library exists yet.
The gap
Productionizing the academic methods. Latency overhead under 5%. Compatibility across vLLM and TensorRT-LLM. NSF NDIF: $9M for inference research. IITP Korea, GuangDong Foundation are funding the adjacent work. Inference optimization engineer roles surged 59% in 2024-2025. The first library to ship benchmarked results on Llama 3.1 70B and Qwen wins the segment.
03 / Deep Tech Bet
One outbreak in a nursing home costs $50k to $150k. A single-dose nasal spray vaccine could prevent it.
What just became possible
A November 2023 bioRxiv paper validated adenoviral-vectored mucosal vaccines that deliver durable protection against SARS-CoV-2 transmission. Mucosal IgA induction lasts past 12 months in elderly cohorts. Spray-dried formulations enable room-temperature storage. The platform extends to influenza and RSV with the same delivery mechanism.
Why now
2 million older adults in 65,600 US long-term care facilities. Senior living market: $1.64T in 2026 to $3T by 2033 at 9.3% CAGR. Average outbreak management cost: $50k to $150k per facility, more for 100+ bed sites. Facilities pay for isolation protocols, surge staffing, liability insurance, and resident refusal of needle-based vaccines.
What you'd build
A single-dose nasal spray vaccine for elderly care facilities, formulated for durable mucosal immunity against respiratory pathogens. Buyer: CNOs and Infection Prevention Directors at Sunrise Senior Living, Brookdale, and the 200+ chains running 100+ bed facilities. Pricing: per-resident-per-year contract.
Who's already moving
ENA Respiratory (INNA-051, Phase 2, immune primer). AuraVax Therapeutics (NanoSTING, NIH R01GM143243). GSK, AstraZeneca, Novavax, Bavarian Nordic (legacy infrastructure plus FluMist). Stanford, Duke-NUS, University of Houston on academic research. No one has a senior-living-specific go-to-market for nasal sprays.
The gap
FDA BLA or Combination Product pathway for sterile aerosol drug-device. Consistent dose delivery and spray geometry. Non-inferiority vs systemic IgG plus superiority on nasal IgA. NIH Project NextGen has $5B+ for pandemic preparedness. CEPI's 100 Days Mission funds spray-dried RNA. BARDA Project NextGen. 5-7 year timeline standard, possibly faster on EUA or Fast Track. The vaccine company plus a senior-living distribution partner wins.
BASF's Interceptor G2 protects 63 million people. The next 100 million is a $1.5B procurement decision still being made.
What just became possible
An August 2025 medRxiv paper confirmed sustained malaria case reduction from Interceptor G2 (chlorfenapyr + alpha-cypermethrin) in regions with high pyrethroid resistance mediated by oxidases, esterases, and GSTs. 20-50% better than standard nets. 84% of new nets shipped to Africa in 2024 are dual-AI. The transition is in motion but not complete.
Why now
Annual dual-AI ITN spend in sub-Saharan Africa: $1.2B to $1.5B by 2026. The Global Fund has committed $50M (2021-2024) to market entry. Gates Foundation underwrote 40% price reductions. The next funding cycle locks in procurement through 2028. 300 million nets need replacement every 2-3 years.
What you'd build
Not another net. The dual-AI ITN market is BASF (Interceptor G2, 35M+ nets shipped), DCT (Royal Guard), Vestergaard (PermaNet Dual), Sumitomo (Olyset Plus), plus PRONet Duo in development. Build the procurement intelligence layer: resistance maps, durability tracking, and net-mix optimization for NMCPs and Global Fund Revolving Facility decisions.
Who's already moving
BASF, DCT, Sumitomo, Vestergaard manufacture. IVCC coordinates. PMI VectorLink Project supports 17 sub-Saharan countries with embedded technical staff. WHO and Africa CDC lead policy. The Bill and Melinda Gates Foundation funds resistance research. The procurement decision-support layer is the open commercial lane.
The gap
Real-time resistance surveillance plus procurement optimization in one product. WHO endorsement of data-driven net-mix decisions. The next $1.5B in procurement spend will route through whoever owns the data layer. PMI Initiative, Global Fund Grant Cycle 8 (2024-2026), and GiveWell ($41M to AMF for DRC) fund the surrounding ecosystem. The window: 18 months before the next procurement cycle locks in vendor relationships.
05 / Watch This Space
Copper's reflectance has blocked high-speed EV battery laser welding for a decade. A new beam-shaping optic just hit 35 meters per minute, pore-free.
What just became possible
A July 2023 Research Square paper demonstrated Multi-Plane Light Conversion (MPLC) beam shaping for copper laser welding at 35 m/min with no porosity. Cailabs commercialized the CANUNDA-HP module, integrated into Precitec laser heads. Tested up to 16 kW with a 30 kW pathway. Battery busbars and connectors weldable at automotive line speeds.
Why now
Automotive laser welding market: $1.25B in 2024 to $5.06B by 2031 at 5-6% CAGR. EV battery is the fastest segment. CATL, LG Energy Solution, Northvolt, TE Connectivity, Amphenol all need this for busbars. Tesla is hiring laser welding manufacturing engineers right now. DOE put $725M into US battery materials processing in 2025, with $43M specifically for vehicle technologies.
What you'd build
An industrial-grade laser welding head with MPLC beam-shaping optics integrated for kW-level copper welding. Or build the optics module that retrofits into existing TRUMPF, IPG, and Coherent welding heads. Sell to Tier-1 EV battery assemblers and connector manufacturers. Pricing: per-head plus license.
Who's already moving
Cailabs (MPLC leader, CANUNDA-HP, Precitec partnership). Syspro Automation (BEAMSHAPE consortium, CDTI grant). TRUMPF (BrightLine, green laser). IPG Photonics (Tesla production lines). Coherent (high-power fiber). NVIDIA Isaac robotics for adjacent automation. The whitespace: a productized MPLC welding head designed for the harsh production-floor environment.
The gap
Thermal management of optics inside a robotic welding head. Optical alignment under vibration and 200°C ambient. Scaling MPLC from lab kilowatts to industrial 16-30 kW. DOE Battery Materials grants ($725M total). The first vendor to ship a production-ready 16 kW MPLC welding head plus warranty wins the EV battery line CapEx.
See you next week.
- Theis