The Investor

The Marks Are In

The write-downs the market whispered about in 2022 are now on the books, set by the people whose job is setting them. T. Rowe Price, Fidelity, and Franklin Templeton have taken Airtable, worth eleven billion dollars in 2021, down at least 60%. DataRobot, six billion dollars in mid-2021, is marked to near zero. Gusto is off about thirty percent. There is a reading where these are rate-driven repricings that reverse the moment sentiment turns, and it is not a crazy reading. It is just not what these look like. These look like customers replacing the software, not repricing it as AI eats the application layer.

An Atlanta real-estate manager pulled Salesforce out entirely and rebuilt the thing on Replit + Claude Code for $100K/year in savings, which is the sort of anecdote that means nothing until five SMBs kill their Salesforce and HubSpot contracts inside six months and do the same. HubSpot's stock is down 75% since early 2025, and its four-day reversal on opt-out AI data collection is the tell: it could not retrofit an AI moat without triggering churn.

Infrastructure Versus Application Layer

The same sector is throwing off opposite outcomes, which is the part worth sitting with. The line is not AI branding. DataRobot sat squarely in AI/ML and still got zeroed. The line is defensibility against foundation models.

Battery's Brandon Gleklen put it bluntly: product-market fit that used to buy a decade of growth is now ephemeral, bookings spike and the foundation models catch up a quarter or two later. Oquirrh's Ron Heinz says software values keep trending down except for companies with very high growth rate or technology hard to replicate.

Why HubSpot Backed Down in Four Days

HubSpot's data revolt is the case worth studying. On July 1 it announced opt-out AI data collection. Four days later it reversed and called the whole thing 'a mistake.' Legacy SaaS cannot bolt on a pooled-data AI moat without detonating its own base, because customers now treat CRM data as a defensible asset and have credible exit options (Attio is winning the defectors on cost and trust). Zoom ran a version of this in 2023, Slack in 2024, HubSpot in 2026, and the pattern is boring enough now to price.

AI did not lower software multiples so much as split them: infrastructure compounds, while applications get rebuilt in a weekend for $100K in savings.

The Deal That Repriced an Asset Class

TeraWulf signed a 20-year, ~$19B lease with Anthropic for a 401MW Kentucky campus — roughly $950M/year of contracted revenue. The stock jumped 14% to $24.05. But the investable signal isn't TeraWulf alone — it's the coordinated 10-14% move across the entire bitcoin-miner cohort in a single session: IREN +13%, Hut 8 +12%, Cipher +11%, Keel +10%. The market re-rated an asset class before lunch.

The same day, TeraWulf sold its $450M Texas mining stake to Fluidstack — as clean a goodbye to crypto as a company can manage. Bloom Energy popped 9% on a separate 15-year Fortune 100 data-center PPA, confirming the thesis extends beyond mining pivots to any entity sitting on interconnected power.

Why This Matters Now

Anthropic is deploying capital at hyperscaler scale while remaining private: $19B Kentucky lease plus a $15B Australian tender. The frontier labs' infrastructure ambitions have outgrown what the cloud alone can supply. The scarce input was never models — it's grid-connected megawatts and shovel-ready sites near cheap power.

The contrast with GPU compute is instructive. Nvidia is now offering neoclouds guaranteed buybacks on unused capacity — setting a price floor for compute while simultaneously signaling demand may not fill supply. Meta's enterprise-cloud pivot, celebrated with a 10% stock pop, is explicitly framed as utilizing potential overbuild. When the monopolist guarantees to buy back its own product, stop paying for scarcity at that layer.

The Risk Framework

Counterparty concentration is the key vulnerability. The miner-pivot cohort's contracted revenue is only as good as Anthropic's balance sheet. McKinsey warns nuclear costs threaten US power ambitions, and interconnection timelines routinely slip. The 20-year duration carries genuine execution risk — construction delivery and sustained compute demand against relentless efficiency gains. Don't comp mechanically; structure for duration risk.

The AI trade's real scarcity isn't compute or models — it's grid-connected power, and the market just repriced every megawatt-holder in a single session.

The Category Is Forming

Mercor hit a $2B+ gross revenue run rate in June — double its pace earlier this year — while remaining free-cash-flow profitable. A three-year-old company that scaled $1M→$500M in 17 months is now compounding past $2B. Its nearest comp, Handshake, crossed $1B. Two independent players at nine-figure-plus scale, both profitable-ish, driven by Fortune 500s fine-tuning their own models rather than renting foundation models wholesale.

The projected TAM is staggering: labs are forecast to spend $100B/year on data by 2030 as high-quality public internet data goes scarce and private datasets become the strategic asset. This is the rare thing an investor gets paid for spotting — a TAM-forming inflection before the category has a consensus name or incumbent.

Why Data Wins Now

Multiple independent signals confirm the bottleneck has shifted:

• Architecture efficiency is squeezing more from less compute — Tencent's Hy3 (295B params, 21B active) matches models 2-5x its size, pushing the constraint onto data quality
• Enterprise adoption lifecycle follows a pattern: frontier closed models for exploration → fine-tuned open-source in production. Decagon already runs 90% on open-source
• Open-weight commoditization erodes model pricing everywhere except reliability — the durable differentiation lives in proprietary training data, not capability

The Mercor Comp

Mercor's $10B valuation set 9 months ago pencils to ~12-16x net revenue (after 60-70% contractor payouts on $2B gross). At 100%+ growth, that looks like a floor, not a ceiling — but the unit economics require scrutiny. This is an expert-marketplace business where the platform's moat is matching quality and network density, not the data itself. The alpha is in sourcing earlier-stage domain-data startups before the category-wide re-rate prices you out — but underwrite net revenue rigorously.

Compute value has been captured — the next Stargate is data, and it's the only whitespace in this cycle still trading below consensus.