AI's true cost
1. Key Themes
Theme 1: AI's True Cost — Headline Prices Are Misleading
The market-consensus view that cheaper AI models save money is flawed. Reliability is the hidden variable that determines total cost of ownership.
"The headlines about frontier AI models cutting prices hide the true cost to customers. And with businesses increasingly watching their AI spending, that illusion can lead to expensive mistakes."
"A cheaper model that fails more often has to re-run the task, adding more tokens to the bill and, more expensively, human labor to get it across the finish line."
Concrete evidence: Claude Opus 4.8 (near the top of the price list) completed a typical business task for ~$2.56 all-in, while Meta's Muse Spark (cheapest per token at $4.25/million) cost $5.83 for the same job — more than double.
Theme 2: AI Model Subsidization Is Unsustainable
The lowest-priced models are being sold below cost, propped up by parent-company subsidies. Anthropic's upcoming IPO will force a market reckoning.
"The lowest-priced AI models are being sold for less than they cost to run—subsidized by the ad revenue, parent-company cash and borrowed money."
"That subsidy will likely face its first real test in the fall, when Anthropic reportedly plans to go public. The market will finally see how much money these models actually make—or lose."
Theme 3: Canadian Pension Funds Doubling Down on US Industrial Real Estate
CPP Investments and Brookfield are acquiring LXP Industrial Trust for $5.2B, continuing the "Canadian model" of direct asset ownership — specifically targeting Sun Belt logistics assets despite recent real estate struggles.
"Direct investments are key elements of the 'Canadian model.' The investment approach, pioneered by the major Canadian public pension funds, is characterized by direct asset ownership and in-house management. This approach has historically generated the strongest returns in the real estate asset class."
"The US industrial sector remains attractive to CPP Investments due to 'structural demand drivers, including domestic manufacturing, evolving global supply chains and population growth across key Sun Belt markets.'"
Theme 4: Independent Sponsors (Fundless PE) Are Gaining Ground
As traditional PE fundraising stalls, deal-by-deal models are emerging as a viable — and increasingly validated — alternative.
"With PE exits and fundraising both stuck in a rut, deal-by-deal models and support from Washington are boosting returns for independent sponsors once seen as failures."
Theme 5: VC Dealmaking Is Recovering — Founder-Friendly Conditions Returning
PitchBook's VC Dealmaking Indicator is moving into startup-friendly territory, driven by valuation step-ups.
"Our dealmaking indicator continues to move into startup-friendly territory, driven mainly by valuation step-ups."
2. Contrarian Perspectives
Contrarian 1: Expensive AI Models Are Actually Cheaper to Operate
Against the prevailing narrative that businesses should default to the lowest-cost AI API, this article argues the opposite: premium-priced frontier models may deliver lower total cost because they fail less frequently.
"More expensive, newer frontier AI models can get a job done for a lower final cost to the customer. Anthropic's Claude Opus 4.8, currently one of the priciest on the market, came out on top."
Evidence: Muse Spark at $4.25/million tokens (cheapest sticker price) cost $5.83 to complete the task; Claude Opus at a much higher per-token rate cost only $2.56. The failure-and-retry loop is the hidden cost multiplier.
Contrarian 2: AI Token Price Cuts Are Not What They Seem — It's Predatory Pricing Funded by Subsidies
Most observers celebrate AI price deflation as a win for enterprise customers. The article argues this is a distortion created by parent-company cross-subsidies, not sustainable unit economics — and the illusion is about to be exposed.
"The lowest-priced AI models are being sold for less than they cost to run—subsidized by the ad revenue, parent-company cash and borrowed money."
Evidence: Anthropic's planned IPO will be the first major transparency event forcing disclosure of actual model economics. xAI's Grok, for instance, is "cheap per token but hasn't built a meaningful business selling to companies."
Contrarian 3: Going "Fundless" in PE Is No Longer a Stigma — It May Be a Strategic Advantage
The traditional VC/PE framework treats not having a committed fund as a failure mode. This article reframes the independent sponsor model as adaptive and potentially superior in the current environment.
"Deal-by-deal models and support from Washington are boosting returns for independent sponsors once seen as failures."
3. Companies Identified
Anthropic Description: AI safety and model company, maker of the Claude model series. Why mentioned: Its Claude Opus 4.8 delivered the lowest all-in task cost in PitchBook's experiment despite being among the most expensive per token; also planning an IPO that will be the first major stress-test of AI model unit economics. Quote: "Anthropic's Claude Opus 4.8, currently one of the priciest on the market, came out on top."
Meta (Muse Spark) Description: Big Tech conglomerate; Muse Spark is its AI model offering. Why mentioned: Used as a cautionary case study — the cheapest model per token produced the highest real-world task cost. Quote: "Meta's Muse Spark, the cheapest on paper at $4.25 per million tokens, cost about $5.83 to finish the same job—more than double."
Google (Gemini 3.1 Pro) Description: Big Tech; Gemini is its frontier AI model. Why mentioned: Ranked just behind Claude Opus in total task cost efficiency. Quote: "Google's Gemini 3.1 Pro came in just behind Opus."
xAI (Grok) Description: Elon Musk's AI company. Why mentioned: Flagged as cheap per token but commercially underdeveloped. Quote: "xAI's Grok is cheap per token but hasn't built a meaningful business selling to companies."
Brookfield Asset Management Description: Global real assets investment manager. Why mentioned: Co-acquirer of LXP Industrial Trust for $5.2B, executing on the "Canadian model" of direct asset ownership. Quote: "Brookfield Asset Management and the Canada Pension Plan Investment Board have entered into an agreement to acquire real estate investment trust LXP Industrial Trust for $5.2 billion."
CPP Investments (Canada Pension Plan Investment Board) Description: Major Canadian public pension fund. Why mentioned: Strategic acquirer of US industrial real estate; cited as pioneer of the "Canadian model"; previously posted a 5% real estate loss but recovering via logistics focus. Quote: "CPP's real estate investments posted a 5% loss over the 2024 fiscal year, which it attributed to losses in the office and retail sectors and to high interest rates. The logistics sector…proved to be an exception."
LXP Industrial Trust Description: US industrial REIT focused on warehouses and distribution in the Sun Belt and Midwest. Why mentioned: $5.2B acquisition target for Brookfield and CPP Investments; case study for industrial real estate demand thesis. Quote: "The real assets investment manager and pension fund manager will become the owners and operators of LXP's warehouses and distribution facilities in the US Sun Belt and Midwest regions."
Etched Description: AI chip developer. Why mentioned: In talks to raise two rounds — one at $10B valuation (led by Sequoia), one at $20B (led by Jane Street). Quote: "Etched, an AI chip developer, is in talks to raise two rounds: one led by Sequoia at a $10 billion valuation, and one led by Jane Street at a $20 billion valuation."
CuspAI Description: UK deep tech startup focused on AI-driven materials discovery. Why mentioned: Raised $450M Series B at $2.6B valuation from Kleiner Perkins and NEA — a notable non-US AI deal. Quote: "CuspAI, a UK deep tech startup focused on materials discovery, raised a $450 million Series B led by Kleiner Perkins and NEA, valuing the company at $2.6 billion."
Sereact Description: Germany-based physical AI for warehouse robotics. Why mentioned: Extended Series B to $116M with Zalando as new investor; signals retail/logistics convergence with AI robotics. Quote: "Sereact, a Germany-based provider of physical AI for warehouse robotics, expanded its Series B to $116 million after Zalando joined the round."
Neo Description: Boston-based AI governance and security platform for enterprises. Why mentioned: Emerged from stealth with $75M Series A (a16z + Bessemer) plus a previously undisclosed $25M seed — signals rising demand for AI security/governance tooling. Quote: "Neo, which helps enterprises secure and govern AI software, emerged from stealth with a $75 million Series A led by Andreessen Horowitz and Bessemer Venture Partners."
Scribe Therapeutics Description: a16z-backed gene editing biotech. Why mentioned: Eyeing $107.2M IPO at ~$242.7M valuation. Quote: "Andreessen Horowitz-backed gene editing biotech Scribe Therapeutics is seeking to raise up to $107.2 million in its IPO, which would give the company a valuation of about $242.7 million."
Jersey Mike's Description: Blackstone-backed sandwich chain. Why mentioned: Looking to raise up to $1.09B in IPO at ~$8B valuation — a major PE exit signal. Quote: "Blackstone-backed sandwich chain Jersey Mike's is looking to raise up to $1.09 billion in its IPO, which could give the company a valuation of around $8 billion."
Partners Group Description: Global private markets asset manager. Why mentioned: Closed fourth direct infrastructure program at over $15B, targeting power generation, AI infrastructure, and energy security. Quote: "Partners Group closed its fourth direct infrastructure program, which targets assets spanning power generation, AI infrastructure and energy security, on over $15 billion."
Tempus Description: Healthcare-focused AI company. Why mentioned: Agreed to acquire Personalis for $1.25B — a significant AI-in-healthcare consolidation move. Quote: "Tempus, a healthcare-focused AI company, agreed to acquire Personalis, which offers personalized testing for cancer management, in a $1.25 billion deal."
Hugging Face Description: Open-source AI model hub and company. Why mentioned: Turned to a Chinese AI model after an American one failed to contain a cyberattack on internal datasets — a geopolitically sensitive anecdote. Quote: "Hugging Face turned to a Chinese AI model after an American one failed to stop a cyber attack. Internal dataset breaches caused by the attack could only be stopped by an outside model."
4. People Identified
Harrison Rolfes Description: Senior Research Analyst, Private Company Coverage at PitchBook. Why mentioned: Author of the featured analysis on AI model true costs (the "Anthropic's priciest tokens do the cheapest work" piece). Quote: Byline credit for the AI cost analysis section.
Jessica Hamlin Description: Senior Funds Columnist at PitchBook. Why mentioned: Author of the LXP Industrial Trust / CPP Investments acquisition piece. Quote: Byline credit for the Brookfield/CPP/LXP section.
Sophie van Oosterom Description: Global head of real estate at CPP Investments; previously from Schroders Capital (joined January 2025). Why mentioned: Articulated the strategic rationale for the LXP deal; this is her second industrial REIT acquisition at CPP after a period of real estate underperformance. Quote: "The US industrial sector remains attractive to CPP Investments due to 'structural demand drivers, including domestic manufacturing, evolving global supply chains and population growth across key Sun Belt markets.'"
Claudia Goldin Description: Nobel laureate economist. Why mentioned: Cited in Side Letters for her argument that college-educated women are forgoing motherhood when partner dependability is uncertain, contributing to falling birth rates in developed economies — a macro demographic risk factor. Quote: "College-educated women are skipping motherhood when a partner's dependability is in doubt, Nobel laureate Claudia Goldin argues, as birth rates continue to fall in developed economies."
5. Operating Insights
Insight 1: Benchmark AI Models on Total Task Cost, Not Token Price
When procuring AI models for business workflows, don't default to the cheapest per-token option. Run head-to-head comparisons that include retry rates and human labor costs to remediate failures. The highest-priced model may yield the lowest total cost.
"AI models are billed by the 'token'…so a cheaper model that fails more often has to re-run the task, adding more tokens to the bill and, more expensively, human labor to get it across the finish line."
Insight 2: PE Buyout Multiples Are Back at 2021 Highs — Underwrite Carefully
Deal discipline is critical: the median US buyout multiple hit 12.5x EV/EBITDA in 2025, matching the peak 2021 level. Operators and acquirers must stress-test their return assumptions against this elevated entry point.
"The median US buyout deal was done at an enterprise-value-to-EBITDA multiple of 12.5x in 2025, matching the record level seen in 2021."
Insight 3: European Defense Is an Emerging Private Capital Opportunity
Government budgets alone can't meet Europe's defense needs, opening a capital gap that PE and VC are being called to fill — a nascent but growing investment vertical.
"'Public spending will not be enough,' said the global co-head of investment banking at Barclays, speaking about the needs of defense companies in Europe. Cash-strapped governments are leaving a gap, and PE and VC must fill it."
6. Overlooked Insights
Overlooked Insight 1: Jane Street as a Lead Investor in an AI Chip Deal
Jane Street — traditionally a quantitative trading firm, not a venture capital investor — is reportedly leading one of Etched's fundraising rounds at a $20B valuation. This is a non-obvious signal of financial-market crossover into deep tech venture, and may indicate that sophisticated trading firms see proprietary AI chip infrastructure as strategically essential to their own operations.
"Etched, an AI chip developer, is in talks to raise two rounds: one led by Sequoia at a $10 billion valuation, and one led by Jane Street at a $20 billion valuation."
Overlooked Insight 2: Private Credit Continuation Vehicles Are Pricing at a Premium to Par
At least five recent private credit secondaries continuation vehicles have priced at a nominal premium to par — a subtle but meaningful signal that LP demand for private credit exposure is strong enough to push prices above face value, potentially compressing future returns in the asset class.
"Pricing details have emerged on several private credit continuation vehicles that closed in the past year, showing that at least five of these secondaries transactions have priced at a nominal premium to par."