Fontlume #6: When milliseconds are revenue

What just became possible

A February 2026 arXiv paper showed 2x inference speedup for diffusion models (image and video) by dynamically adjusting timesteps and rectifying caching errors. No quality loss. Drop-in replacement for the standard inference pipeline. Works on NVIDIA H100, AMD MI300X, AWS Inferentia with kernel-level abstractions.

Why now

Runway raised $315M in February 2026 (total $540M+, $3B+ valuation). Pika at $470-700M. Luma raised $900M Series C in November 2025. Synthesia at $4B valuation. They are all burning cloud GPU spend. ML infrastructure engineer demand grew 45% YoY in 2025 with median $185k-$285k. The buy decision will outpace the build decision.

What you'd build

A drop-in inference optimization library that cuts diffusion model GPU spend by 50% with temporal consistency guarantees. Sell to Series A/B video startups running 500+ GPU hours per day. AI video generation market: $18.6B in 2026 to $246B by 2034 at 36% CAGR. Pricing: percentage of cloud spend saved.

Who's already moving

NVIDIA TensorRT (Adobe Firefly uses it, FP8 on Hopper). Hugging Face Diffusers (open-source). BentoML (deployment). Together AI (optimized stack). All optimized for LLMs or static CV. None handle dynamic shape and temporal consistency for high-fidelity video without manual tuning. That is the wedge.

The gap

Kernel-level work that survives the next DiT or Flow Matching architecture. Hardware-agnostic across NVIDIA, AMD, AWS Inferentia. Temporal consistency proofs for video. US Technology Modernization Fund and NSF Advancements in AI for Science fund pieces. The play: open-source core, paid runtime, integration support. Land 3 of the top 10 video gen startups and become the dependency.

What just became possible

An October 2025 arXiv paper proposed a dual-gate governance framework requiring machine-checkable evidence on every LLM claim, with deterministic verification against structured and unstructured sources. SMT solvers plus temporal logic plus an evidence graph. Recent work shows 96%+ reduction in hallucinations for code and logical tasks. The architecture is now reproducible.

Why now

Quantitative hedge funds and prop trading firms manage over $5.3T in assets. Global quant fund market: $16T in 2024 going to $31.4T by 2032 at 10% CAGR. Bloomberg, FactSet, Refinitiv have no verification layer. Early AI adopters report 15% improvement in signal accuracy from alternative data. The buyer is ready and the willingness to pay is in the millions per seat.

What you'd build

A verifiable autonomous research agent API. Outputs only machine-checked claims with attached evidence links and cryptographic audit trails. Replaces junior analysts who manually cross-reference earnings transcripts against SEC filings. Sell to top-20 quant funds first. Pricing: per-seat enterprise contract starting at six figures.

Who's already moving

TauricResearch released TradingAgents as an open-source multi-agent framework getting hedge fund attention. Confident AI's DeepEval handles LLM evaluation but is not a full agent. Artificial Genius does deterministic non-generative fine-tuning. Bloomberg, FactSet, and Refinitiv lack native LLM verification. The whitespace: turnkey API with provable zero-hallucination SLA.

The gap

Building the symbolic verification engine at financial-document scale. Onboarding source providers for the evidence graph. Cryptographic audit trail design that satisfies compliance teams. Meta Llama Research Grant offers $50k. NSF AI Institute funds verifiable AI research at $20M+ scale. Talent: LLM data scientists with formal methods. Salary range starts at $200k and climbs fast.

What just became possible

An April 2026 arXiv paper showed real-time constrained control with guaranteed safety and recursive feasibility without the cost of online optimization. Control Barrier Functions plus Poisson Safety Functions plus pre-computed control laws. Hard safety constraints without the model predictive control compute burden.

Why now

Tier-1 automotive suppliers are spending on NVIDIA Orin and Thor hardware partly to brute-force MPC latency. TÜV and UL safety validation runs $100k to $500k per vehicle program. The autonomous driving software market is $2.36B in 2026 going to $7B by 2035 at 13.3% CAGR. UNECE and ISO standards are pushing toward formal safety proofs and pre-computed laws meet that bar more cleanly than online optimization.

What you'd build

A drop-in safety layer SDK for autonomous robotics and industrial automation. Pre-computed control laws with provable safety guarantees, certified to ISO 26262 ASIL-D. Buyer is the safety certification engineer at Aptiv, ZF, Continental, Bosch, Magna. Sell as a license plus per-vehicle royalty.

Who's already moving

FORT Robotics (industrial safety platform, 300+ customers). Iris Automation (drone collision avoidance). Polymath Robotics (industrial vehicle autonomy stack). NVIDIA DRIVE owns the hardware-tied solution. TÜV SÜD and UL Solutions own the certification. No one offers pre-computed, provably safe control as a drop-in SDK.

The gap

Productizing the math for high-dimensional non-linear systems. ISO 26262 ASIL-D certification (the hard one). Control theorists who write production code, or software teams who can read safety proofs. NSF NRI 3.0 up to $1.5M per project. DOT SBIR $200k to $1M. DHS C-UAS at $250M. The first SDK to ship with ASIL-D paperwork wins the Tier-1 RFP cycle.

What just became possible

An April 2023 medRxiv evaluation validated the Xpert MTB/XDR for rapid detection of isoniazid, fluoroquinolone, and second-line drug resistance in under 2 hours. WHO prequalified. Cepheid's Global Access Program offers cartridges at $7.97 to eligible countries. Solar-powered mobile clinics with cellular backhaul are demonstrably feasible (Samsung's South Africa deployments, PIH Lesotho).

Why now

The Global Fund Grant Cycle 8 (2024-2026) allocated over $13B for HIV, TB, malaria. 12% of TB budgets are earmarked for public-private mix and innovative diagnostics. WHO Europe's PASS to End TB put $50M into advanced diagnostics. India alone has 7,767 rapid molecular testing facilities, with rural deployment as the gap. The cartridges are subsidized. The kiosk infrastructure is not.

What you'd build

A turnkey solar-powered kiosk pre-loaded with Xpert MTB/XDR cartridges, automated digital reporting, and cellular backhaul to national TB databases. Solar PV cost for rural Uganda clinics is $10k to $20k per site. Bundle the kiosk, the cartridges, and the maintenance contract. TB diagnostics market: $2.53B in 2026 to $4.14B by 2034.

Who's already moving

Cepheid sells GeneXpert Edge (battery) but not the integrated kiosk. Molbio (Truenat) is the chip-based competitor in India and Africa. Tulane's LIT (lab-in-tube) is in trials. Partners In Health runs solar mobile health trucks in Lesotho. Samsung built mobile solar clinics in South Africa. No one bundles the molecular cartridge, kiosk, and digital reporting as a single SKU yet.

The gap

IP licensing with Cepheid for the bundled product. CE/WHO re-certification of the integrated kiosk. Ruggedization against 45°C heat, dust, and vandalism. Local maintenance partnerships. The Global Fund has emergency funding for mobile units. Stop TB Partnership runs the Digital Health Technology Hub. Whoever proves operational economics in one country at scale gets cloned across 20 more.

What just became possible

A June 2024 arXiv paper presented recursive polynomial optimization for collision-avoidance maneuvers under Keplerian, J2, and CR3BP dynamics, computing optimal trajectories in under one second. Validated through peer review in IEEE and AIAA. ESA and NASA have open data standards for conjunction messages, so integration is well-defined.

Why now

Starlink performed 144,404 collision-avoidance maneuvers between December 2024 and May 2025. Constellation density is doubling. SpaceX is hiring Starlink astrodynamics engineers and flight dynamics engineers. Amazon Kuiper is ramping up. OneWeb scaling. Over 100,000 satellites projected in LEO by 2030. The conjunction assessment bottleneck is now a daily operational crisis.

What you'd build

A real-time collision avoidance API for satellite operators. Sub-second optimal maneuver computation. Plug into existing ground-station software. Sell to commercial constellation operators outside the SpaceX in-house lane. LEO satellite market: $7.23B to $32.59B in 2026, scaling to $20.69B to $54.62B by 2030. Collision avoidance software alone is roughly 24% of that.

Who's already moving

AGI/Ansys STK is the legacy. SpaceX Stargaze is in-house and not licensed. Lockheed Martin has AI-based navigation in their bus. LeoLabs (Collision Avoidance API, used by Starlink and OneWeb). GMV is developing next-gen for LEO constellations. OKAPI:Orbits and Neuraspace run ML approaches. Space Guardian (ESA-backed, €72k pre-seed) is the newest. Crowded, but no one has the sub-1s polynomial method productized.

The gap

Trust and liability are the real barriers. Operators fear automated maneuver decisions because the cost of false positives is fuel-burn and the legal liability for an automated collision is unmapped. ESA ARTES (€1M to €5M), NASA SBIR ($150k to $1M), ESA CASSINI fund the technical work. Win one mid-tier constellation (OneWeb, Iridium-NEXT successor, a defense contract) and the rest follow.

See you next week.

- Theis