The Investor

Two unaffiliated activist shops publishing on the same $69.6B company in one week is rare consensus — and the Bloom Energy thesis is forensic, not narrative. Hunterbrook and Crossroads converge on a scandium supply wall, 74% related-party revenue, and revenue booked on uninvoiced sales — atop a fourth CFO in two years. Related-party concentration plus unbilled recognition plus accounting-officer churn is the textbook restatement precursor. Short cycles at this cap play out fast; exposure decisions can't wait for the quarter.

The template transfers to private diligence: run related-party revenue, invoice timing, and CFO tenure across every hardware and energy pipeline company with steep growth projections. The week's executive tape reinforces it — Fiserv lost CEO and President within a month; Angi's CAO exited in 3 months at a company that cycled 5 CEOs and 5 CFOs in a decade; Veraxa Biotech is down ~80% one month post-SPAC with its CFO leaving 'with immediate effect.' Accounting-officer flight is the cheapest fraud signal available; remove de-SPAC as a base-case exit in any liquidity plan.

The structural half is Intel: the US government converted $9B in grants into a 10% stake — now its largest shareholder — while pressuring Apple to use Intel fabs. State capitalism has arrived in semiconductors: every domestic semi and foundry thesis now competes against an incumbent with a government demand floor and political air cover, reshaping win rates for challenger fabs and packaging plays and adding a policy-overhang discount to rivals for the same anchor customers.

When two independent shorts and a CFO carousel converge on one balance sheet, repricing beats your quarterly review cycle — act on exposure first, refine the thesis second.

The contradiction nobody has priced: AI's input cost (GPU-hours) and output price (tokens) are drifting apart. Silicon Data's one-year H100 index bottomed near $1.70/hr last October, sits at $2.35/hr now, spot up 10% YTD. Meanwhile Grok 4.5 ships at $2/$6 per million tokens and GPT-5.6 Luna at $1/$6, both flagged 'highly subsidised for now,' which is a phrase that finances itself off a balance sheet rather than a P&L. The wedge cannot widen forever. Whichever side snaps first sets your infra and app-layer marks, and that is the whole trade.

Demand looks corroborated by the buyers who actually write cheques. JPMorgan is deploying SambaNova's SN40/SN50 on-prem behind a $1B Series F at $11B. Positron is in talks for roughly $750M at about $5B as a direct Nvidia inference challenger. Regulated enterprises paying for owned inference is durable demand. It is the opposite of a capex cliff, or at least the opposite of the version people worry about.

The caveat is in the fine print. SpaceX's S1 discloses infrastructure deals with a 3-year tenor but 90-day cancellation rights. Prices rise, yet the most sophisticated buyers refuse to lock duration. That means duration risk is migrating from buyer to provider, quietly, one contract at a time. Neocloud 'contracted backlog' is now closer to 90-day optionality than a bond. Model the revenue on utilization, not contract length.

Cross-check the tape and the two stories diverge. Memory names are up +200% to +700% YTD, and Micron's margins now exceed the 2018 peak that preceded a stock decline, which is to say the market is pricing zero mean reversion through 2028. The GPU rebound is real demand data; the memory tape is scarcity-rent extrapolation. Treat them differently.

Compute demand is proven; compute contracts are not — buy the utilization story, discount the backlog story.

The pattern under-priced in this week's seven-billion-scale financings is dull enough that nobody bothered to write it down: capital is funding companies to stop renting frontier APIs. Prime Intellect took $130M in a Series A at a $1B valuation from Radical so that firms like Ramp can train their own models, which is a strange thing to fund unless you think the renting itself is the problem. A credible security operator is now telling people to run downloaded, locally-runnable models for resilience. China's telemetry-as-backdoor posture is pushing regulated buyers toward sovereign inference. Three independent sources, one direction. This is probably reading a demand curve into three data points that happen to rhyme, but on-prem and owned models look like a curve forming before consensus catches it.

Ollama resets the comp sheet, and here the numbers do the arguing: 8.9M monthly developers, twice January's figure, 85% of the Fortune 500, a million installs a week, built by 14 employees on $88M total raised after a $65M Series B from Theory and Benchmark. Distribution built before monetization is the batch's scarcest asset. Any dev-infra position burning five times that for comparable reach earns a hard conversation, or at least a question about whether reach was ever the constraint.

Also worth logging: Norm Ai's $120M Series C at $1.2B from Khosla pairs AI agents with an affiliated AI-native law firm and supervisory agents — an 'agent plus accountability' structure that is repeatable enough to show up in healthcare and finance next. Lovable's $300M at $13.2B is the other kind of data point, double its roughly $6.6B December mark in seven months, and that is the cycle's froth marker. Use it as the ceiling on app-layer step-ups, not the baseline. OpenAI sunsetting its Atlas browser under a year in confirms app-layer surfaces without workflow lock-in are features, not destinations.

Governance is the diligence gate this time, not the roadmap slide. Agents running for hours, touching financial models and filesystems, turn permissions, audit, and rollback into pass/fail tests. The version where any of that stays a feature request does not survive contact with a compliance team.

The scarcest assets this cycle are capital-efficient distribution and the infrastructure that lets enterprises own their models — everything priced on renting frontier access is melting ice.

The genuinely new fact: xAI trained Grok 4.5 on 'entire data from Cursor' — a platform's user data harvested to build a competing model. That torches the assumption that data-network effects protect app-layer incumbents by default. Data rights are now a first-class diligence line: any portfolio company exposing user code or workflows to a third-party model provider without airtight use clauses is building its moat on sand. Run the audit before the next term sheet.

Anthropic is squeezed on two fronts. Price: Opus 4.8 lists at $5/$25 against Grok 4.5's $2/$6 and GPT-5.6 Luna's $1/$6 — a 3-5x premium. Capability: Meta's Muse Spark 1.1 beats Opus on JobBench (54.7 vs 48.4); Claude Fable 5 trails GPT-5.6 Sol on coding (77.2 vs 80.0). The hidden third front is token efficiency: Grok completes SWE Bench Pro tasks in 15,954 output tokens vs Opus's 67,020 — 4.2x — so the real cost-per-task gap is wider than sticker prices imply. Don't confuse a pricing loss with a franchise loss: switching costs and the safety brand can sustain revenue while pricing power erodes — but the premium is structurally exposed into any 2026 listing.

As models converge, value accrues to distribution and orchestration. Cursor is the kingmaker every lab courts — xAI trained alongside it, and it's doubling usage allowances at launch — a neutral IDE layer collecting rent from competing model makers. Below it, the harness layer (Pi: 1.2–2.08x cost savings) and open-source routing (Plano cut one agent's usage 2x with zero code changes) show mechanics going free while durable advantage concentrates in preference data, observability, and governance.

When platform data can train your competitor, data rights — not benchmarks — become the first question in every AI diligence.