Three sessions.
One vertical.

Settled: photonic computing offers theoretical advantages in matrix multiplication operations relevant to AI inference, principally lower energy per operation at scale. Contested: whether any current photonic chip achieves these advantages outside laboratory conditions at commercially relevant precision levels. Unknown: scalability of photonic interconnects beyond single-chip demonstrations; yield rates in commercial silicon photonics fabrication.

Claims analysis: the company's benchmark figures use INT8 precision. Competitive AI accelerators (NVIDIA H100, AMD MI300X) are benchmarked at FP16/BF16. Comparing INT8 photonic throughput to FP16 GPU throughput produces an inflated efficiency ratio. The honest comparison requires matched precision benchmarking that has not been published. This is not a minor methodology difference — it changes the headline efficiency claim by approximately 4×.

“The precision mismatch is the core diligence finding, but there's a second problem: the demo environment was a single-tenant rack with no PCIe bandwidth constraint. Real inference workloads are memory-bandwidth-bound at the host interface. The photonic advantage disappears entirely when host memory bandwidth becomes the bottleneck. No published study has characterised photonic inference throughput in a memory-bandwidth-constrained production environment.”

Confidence: Contested that commercial claims are supported by published evidence. Gap on memory-bandwidth-constrained production performance. Diligence recommendation: require matched-precision third-party benchmarking in a production-representative environment before term sheet.

Settled: the clinical documentation software market (EHR-adjacent) is approximately $8–12B annually in 2024. Contested: the AI-enabled penetration rate and pricing premium assumptions that drive the $47B figure — vendor projections in this space have historically overestimated AI adoption rates by 3–5 years. Unknown: whether reimbursement policy will evolve to support AI documentation tools or whether payers will treat AI-generated notes as equivalent to human-generated notes.

Projection methodology: the company's TAM was built bottom-up from per-physician pricing × physician count × penetration rate. The penetration rate assumption (68% by 2029) has no historical analogue in healthcare software. The fastest healthcare software category adoption on record is electronic prescribing, which reached 68% penetration in 11 years from regulatory mandate. AI documentation has no regulatory mandate.

“The ePrescribing comparison understates AI documentation momentum because EHR fatigue is a documented physician crisis that creates demand-pull without regulatory mandate. The 68% penetration assumption may be optimistic, but the directional case for rapid adoption is stronger than the historical analogy suggests. The question is whether the price point holds — the TAM assumption uses $180/physician/month, which is 3× current EHR add-on pricing.”

Diligence focus: the price point assumption is more fragile than the penetration assumption. Model sensitivity: at $60/physician/month (current add-on comp), TAM is $16B at 68% penetration — still a large market but changes the return model. Request bottoms-up pricing analysis with customer willingness-to-pay data before closing.

Settled: LBO covenant breach rates correlate strongly with EBITDA-to-interest coverage ratios and are predictable from public capital markets data. Contested: the degree to which current credit market conditions (elevated base rates, tightened spreads) will propagate into covenant stress for 2021–2022 vintage deals. Unknown: whether Fed rate trajectory will provide sufficient relief to extend runway for deals with 2025–2026 covenant reset dates.

Historical base rates: in the 2008–2009 cycle, LBO covenant breach rates for deals with coverage ratios below 1.5× reached 34% within 18 months of the rate shock onset. Current portfolio company has 1.6× coverage. Base rate breach probability at this coverage level in comparable cycles: approximately 22%. If EBITDA deteriorates 10% (within forecast range), coverage drops to 1.44× and breach probability rises to 31%.

“The 2008–2009 base rates are the wrong comparison class. That cycle involved a credit freeze that eliminated refinancing options entirely. Current credit markets remain open to refinancing for investment-grade-adjacent credits. The relevant base rate is 2015–2016 energy sector stress, where companies with 1.5–1.7× coverage breached at approximately 14% with active management. The portfolio company has refinancing options that the 2008 comps did not.”

Confidence: 22–31% breach probability range under base and downside EBITDA scenarios using 2008 comps; 14–20% using 2015 energy comps. Recommend: run both scenarios in board materials and present as a range rather than a point estimate. Active refinancing pursuit is warranted regardless of which base rate is correct.