Fontlume #5: Three races where the wrong differentiator still wins

What just became possible

A January 2024 Research Square paper presented rationally engineered compact BlCas9 variants with enhanced editing activity and expanded PAM compatibility, sized for single-AAV packaging. SaCas9, Nme2Cas9, and SpaCas9 orthologs have been refined the same way. Off-target rates approach SpCas9 levels.

Why now

Dual-AAV delivery (split-intein, trans-splicing) is expensive and reconstitution efficiency is low. Intellia and CRISPR Therapeutics own the proprietary compact-nuclease platforms with high licensing fees. The AAV gene therapy market is $5.4B in 2026 going to $112B by 2035 at 40% CAGR. The packaging-limit constraint is now the binding constraint on every in vivo program.

What you'd build

An IP-clear library of validated compact, high-fidelity Cas9 variants licensed to clinical-stage gene therapy biotechs working on DMD, CFTR, and complex multi-gene disorders. Or run it as a CRO with the variants paired to AAV manufacturing. AAV vector manufacturing alone: $1.66B in 2026 to $6.25B by 2035.

Who's already moving

Intellia and CRISPR Therapeutics own the gold-standard compact platforms. Scribe Therapeutics (CasX, partnered with Biogen). Dyno Therapeutics does AI-driven capsid engineering with Novartis and Roche. Passage Bio ($379.5M), Neurogene ($383.5M), Akouos, AAVantgarde all need compact variants but don't own them. 953+ contract gene therapy AAV jobs open right now.

The gap

Empirical screening across thousands of variants is expensive but cheap relative to losing a clinical program. NIH SCGE put $190M into somatic cell genome editing (2018-2024). UC Davis got $9M for safe editing tools. The IP whitespace is real: compact-Cas9 patents are less crowded than SpCas9 patents. First-mover with a 50-variant library in clinical-grade vectors wins the licensing market.

What just became possible

A June 2025 medRxiv paper validated real-time AI prediction of sepsis risk before triage, accelerating antibiotic administration. Bayesian Health's TREWS, deployed at Johns Hopkins and Cleveland Clinic, hit 17 to 18% sepsis mortality reduction and a 10x reduction in false positives versus Epic Sepsis Response. The capability is proven.

Why now

CMS SEP-1 compliance ties to Value-Based Purchasing penalties starting in 2026. 4.7 million US ED visits per year involve sepsis. 80% of hospitals already use vendor-supplied AI modules. 66% of physicians use AI tools, up 78% from 2023. The buyer's calendar is dictated by CMS, and the budget exists.

What you'd build

A real-time sepsis risk scoring API that integrates with ED triage workflows, ships with XAI dashboards and audit trails, and gives CMS SEP-1 reporting out of the box. Buyers are CMIOs and CNOs at IDNs with over 500 ED visits per day. Sepsis diagnostics market: $885.7M in 2026 to $2.3B by 2035 at 11% CAGR.

Who's already moving

Bayesian Health (TREWS, Johns Hopkins, NSF-backed). Dascena (InSight, FDA-cleared, 48h prediction). UC San Diego's COMPOSER (in-house, 150+ variables, 17% mortality reduction). Epic Sepsis Response and Cerner PowerChart are the incumbents everyone is unhappy with. Third-party APIs have a real wedge because the EHR vendors are bad at this.

The gap

Liability acceptance from CMIOs is the soft barrier. Alert fatigue mitigation. CMS SEP-1 alignment baked into the workflow, not a separate report. NSF SCH offers $1.2M over four years. NIH R18 examines AI impact on safety. NSF SBIR up to $2M. The talent moat is real but the deployments are where the value lives. Whoever lands the next 20-IDN reference list owns the segment.

What just became possible

A May 2025 medRxiv paper validated rapid, simultaneous detection of hypervirulent, carbapenem-resistant, and convergent K. pneumoniae pathotypes with 100% specificity using a single multiplex PCR. 577 clinical isolates, CV under 0.1%. Compared to 48-72 hour phenotypic culture, this is hours instead of days for ICU patients.

Why now

BioFire FilmArray dominates syndromic panels but doesn't have full carbapenemase + hypervirulence coverage. Cepheid's Xpert Carba-R covers resistance genes only. The May 2025 paper closes the assay-design gap. CLIA certification and CPT codes are the remaining barriers, and both have established pathways. NIAID/DMID has rolling $5M-per-project awards through 2028 for exactly this.

What you'd build

A CLIA-certified rapid multiplex PCR kit for K. pneumoniae detection across hypervirulent, carbapenem-resistant, and convergent strains. Sell to clinical lab directors at academic medical centers with high MDR Gram-negative caseloads. Klebsiella testing market: $1.2B in 2024 to $2.4B by 2032. Multiplex PCR overall: $1.25B to $3.43B over the same period.

Who's already moving

Cepheid (Xpert Carba-R, dominant). BioMérieux (BioFire FilmArray, $3.5B revenue 2024). Luminex (xMAP bead-based multiplex). Nidaanvik (India, early-stage). Huwel Lifesciences (India, ICMR-validated). Molbio (India, EnViro-Q). The whitespace: hvKp-specific claims with the convergent strain panel. India is a serious manufacturing base for this.

The gap

CLIA high-complexity certification. CPT code negotiation. Convincing US labs to buy a standalone panel rather than waiting for FilmArray to add it. ICMR funds an integrated Indian genomic surveillance program for hvKp. CARB-X and BARDA are aligned funders. Antibody discovery scientists earn $62k to $190k; assay scientists are cheaper. The first FDA-cleared hvKp-specific kit wins.

What just became possible

A February 2023 bioRxiv evaluation of PermaNet Dual (deltamethrin-chlorfenapyr on a soft polyester fiber) showed maintained efficacy against pyrethroid-resistant vectors after repeated washing. Three-year field durability. WHO prequalified. Vestergaard is scaling Nigeria production to 10 million nets per year by 2026.

Why now

The dual-AI net transition (80% of nets in 2023) accelerated faster than anyone planned. The Global Fund put $50M in 2021-2024 to support market entry. Gates Foundation underwrites price reductions through Revolving Facility volume guarantees. Vestergaard is building a Nigerian manufacturing hub (600 new jobs by 2026) for PermaNet Dual specifically.

What you'd build

Not a competing net. The dual-AI net market is a duopoly (Vestergaard + BASF + DCT with smaller share). Build the procurement-grade durability and resistance surveillance tooling that lets NMCPs prove which net is working in which district. NMCPs spend over $1B annually. Whoever sells the data that drives net-mix decisions earns a slice of every procurement cycle.

Who's already moving

Vestergaard (PermaNet Dual, market leader). BASF (Interceptor G2, Global Fund volume guarantees). DCT (Royal Guard, niche). No early-stage startups in nets. On surveillance: IVCC supports product development. Target Malaria runs vector genetics. WHO and PMI fund coordination. Commercial surveillance-as-a-service is the open lane.

The gap

Local in-Africa manufacturing capacity is being built. Vestergaard's Nigeria expansion is a signal that vertical-integration plus African production is the new bar. For a non-net startup, the gap is data: durability, behavior change, resistance monitoring, all priced as a service to NMCPs and the Global Fund Revolving Facility. The Bill and Melinda Gates Foundation funds the adjacent surveillance work.

What just became possible

A February 2025 arXiv paper applied speculative decoding to auto-regressive TTS, hitting sub-50ms inference without quality loss. Streaming architectures (Dia2, Chatterbox-Turbo) start generating audio before the full text arrives. Quantized models run on modest hardware. The latency war is effectively over: 75 to 90ms is now table stakes.

Why now

Voice AI got $2.1B in VC in 2024, up 7x from $315M in 2022. 80% of businesses plan to integrate AI voice into customer service by 2026. Contact centers are projecting 39 billion calls by 2029. Cartesia Sonic-3, Deepgram Aura-2, ElevenLabs Flash, Inworld Realtime, Rime Mist v2 all sit in the 75 to 100ms band. Latency is no longer the moat.

What you'd build

Not another low-latency English-only TTS API. The next moat is multilingual prosody at sub-100ms, with on-prem and edge deployment for regulated industries (healthcare, finance, government). AI voice agents: $22B in 2026 to $47.5B by 2034 at 34.8% CAGR. Customer support automation alone is 44% of revenue share.

Who's already moving

Cartesia (90ms), Deepgram (90ms optimized), ElevenLabs (75ms), Inworld (sub-250ms P90), Rime (sub-100ms on-prem). All well-funded. Hugging Face, BentoML, AssemblyAI own pieces of the ecosystem. The on-prem multilingual niche is open for the team that can deliver natural prosody across 30+ languages on a single GPU.

The gap

High-quality, diverse, consented voice datasets are scarce. Multilingual prosody, accents, and emotional range are still uneven. NSF Translation to Practice, DOE AI for Scientific Research ($68M) fund adjacent work. ML engineers in this space earn $185k to $285k. Build with on-prem deployment as the default, not the upsell.

See you next week.

- Theis