The Economics of AI Across Three Layers:
• The hardware layer: generational transfer in free cash flow is underway
• The hyperscalers: compute is becoming an asset class, not a sunk cost.
• The labs: the first profitable frontier lab is about to reprice the whole stack
This week printed its first negative free cash flow since it went public.
signed its fourth chip deal in nine months.
all but confirmed it will rent out compute.
So we are stepping back to map the economics of the AI buildout and highlight where we are finding opportunities.
Our lens is simple: follow the cash through the hardware layer, the hyperscalers, and the labs.
The picture is far more constructive than the “capex-is-out-of-control” headlines suggest.
1. The Hardware Layer: A Generational Transfer in Free Cash Flow
Combined 2026 hyperscaler capex has reached ~$725B, up ~77% from ~$410B in 2025, and our work points to more than $1 trillion in 2027.
That spending thus far has been coming from operating cash flow, but things are changing.
Alphabet just posted its first negative FCF since its 2004 IPO.
At the same time, this is landing as record profitability one layer down, where memory pricing is up a cumulative 435% and ’s gross margins are nearing 86%. Compute alone is a ~$380B line this year, roughly double 2025.
The read-through is broadening well beyond .
’s data-center revenue grew 59% (its best in 15 years) and ’ data-center sales are set to double past $3B, while custom silicon pulls a much wider ecosystem (memory, CPUs, networking, packaging) along with it.
When the hyperscalers guide capex up, they guide demand up for all of them. We continue to see the hardware layer as the best risk-reward in the sector.
2. The Hyperscalers: The Compute-Resale Opportunity
The new bear case is that “compute scarcity is over,” pointing to two of the biggest buyers: SpaceX and Meta turning into sellers.
The deal economics say otherwise.
On our math, a gigawatt costing ~$30B to build throws off ~$14.5B of net income a year at conservative rental rates, a ~2-year payback.
Meta sits on ~7GW, doubling to ~14GW, at an estimated 60-70% utilization; monetizing ~5GW of excess could add ~$70B of income. That is not overbuilding; it is compute becoming an asset class.
The depreciation scare fails the same test: four-year-old GPUs still command rising rental rates, so the “worth zero in year five” assumption behind the bear case doesn’t survive contact with the resale market.
And demand is contracted.
More than $2T of backlog sits across the big four clouds. Alphabet is the cleanest expression: Cloud grew 82% to $24.8B at a 36% margin on a $514B backlog, still short of capacity even after guiding capex to ~$200B, while turning its own TPUs into a revenue stream with Anthropic and Meta.
The clearest proof landed this week. Anthropic contracted more than 11GW of compute across four deals, and three of the four are with hyperscalers: Google, and Azure.
For the hyperscalers, renting compute to the labs isn’t a sign of a glut. It’s the second monetization engine coming online.
With capex now outrunning cash flow, funded increasingly by debt, equity and off-balance-sheet leases, access to capital and credit quality is itself becoming the moat.
The Labs: The P&L Behind the Anthropic IPO
The labs are where demand originates and where the “does it pay?” question gets answered.
Anthropic’s run rate has gone from ~$9B at end-2025 to more than $47B by mid-2026, roughly 80x in a year.
Our token-economics work explains why its expected October listing reprices the whole cycle: inference cost has fallen ~40x since early 2024 while revenue per token fell only ~9x, swinging quarterly gross profit from -$55M to an estimated $1B+ by 3Q26 — the first profitable frontier lab.
Anthropic’s four chip deals across Google, AWS, /Nvidia and total more than 11 gigawatts, with the vendors funding their own buyer.
The economics need only a sliver: at list pricing, monetizing under 1% of a single 2GW deployment supports ~$30B of revenue, and our enterprise work shows token consumption still early on its S-curve. Anthropic economics become the guidepost for the entire stack.
Bottom Line
Where we’re finding opportunities. Follow the cash and the same answer surfaces: the hardware layer offers the cleanest risk-reward as free cash flow shifts from buyers to suppliers.
The market keeps hunting for the peak. The economics keep pointing higher-for-longer.
