How can an AI portfolio remain up approximately 80% for the year and still suffer a forced sale?
Borrowing as much as $4 for every $1 invested made volatility existential. That is the clearest lesson from a market where AI demand is no longer the central question. The harder question is whether that demand produces durable cash returns before leverage, depreciation, customer concentration, or valuation forces the issue.
AI-related private business investment reached a record $1.5 trillion annualized rate in Q2 2026, up 25% year over year. Direct AI investment is estimated to represent 25% to 33% of recent US GDP growth. The boom has reached measurable economic scale, but spending is not the same as return.
Leverage Forced the Sale
Leopold Aschenbrenner’s Situational Awareness fund shows why the distinction matters.
Its bottleneck strategy produced extraordinary gains, including a reported 14 times appreciation in Anthropic, from a $61.5 billion valuation to nearly $1 trillion. The industry thesis did not fail first. The financing did.
Leopold reportedly borrowed as much as $4 for every $1 invested. When several overlapping positions declined together, a successful portfolio became a liquidity problem. Rival funds reportedly sold and shorted those holdings once the distress became visible.
Citadel ultimately purchased the debt-financed portion of the public portfolio, likely at a large discount according to the speaker. Situational Awareness retained its client-funded public holdings and private investments, removed leverage, and reportedly remained up approximately 80% for the year.
Concentration was not necessarily the mistake. AI infrastructure was not necessarily the mistake. The mistake was allowing leverage to decide when the assets had to be sold.
Nebius Proves Scarcity, Not Returns
Nebius offers substantial evidence of demand. The company reportedly grew revenue 684% year over year and has more than $46 billion of contracted backlog. Nvidia owns 9.3%. Its contracts are worth up to $19.4 billion with Microsoft and up to $27 billion with Meta.
One author bought $NBIS at $232.49, citing a claimed 20% increase in AWS GPU rental prices and management commentary that four customers are competing for one GPU. That suggests scarcity remains real.
The valuation case is less disciplined. One model projects revenue rising from approximately $3.3 billion in 2026 to $26.9 billion in 2030, with EBITDA reaching $16.4 billion. Applying 15 times EBITDA produces an approximately $250 billion valuation and an $850 per-share target, alongside a second base case near $750. Yet the same text expects an approximately $500 billion market capitalization.
Those outcomes require materially different assumptions. Every bullish path also depends on exceptional growth continuing through 2030.
Michael Burry’s reported short is framed as his Nvidia depreciation thesis “wearing a different mask.” If GPUs have shorter economic lives than hyperscaler depreciation schedules imply, a leveraged GPU renter could suffer more than Nvidia.
Jensen Huang’s desire for independent neoclouds to counter Google TPUs and Amazon Trainium may strengthen Nebius strategically. It does not remove depreciation, leverage, customer concentration, or the need to earn an adequate return on capital.
Cloudflare Has Distribution Before Monetization
Cloudflare’s core business delivered real operating evidence. Q2 revenue was $696.1 million, up 36% year over year. Current RPO grew 35%, while customers spending at least $100,000 annually reached 4,698, up 27%. Full-year revenue guidance increased to $2.864 billion to $2.870 billion.
The costs were equally real. Non-GAAP gross margin declined from 76.3% to 73.1%. A roughly 20% workforce reduction produced $150.7 million of restructuring charges. GAAP net loss widened to $170.0 million, although one analysis estimates approximately $18 million excluding restructuring. Free cash flow was $56.4 million, an 8% margin.
Cloudflare’s AI opportunity remains more strategic than financial. Management declined to quantify Workers or Workers AI revenue, even after adding roughly 2 million developers in Q2. Its agentic-commerce products position the company around identity and payments, but none has disclosed revenue.
Processing roughly half a billion requests per second provides distribution. It does not, by itself, prove monetization.
Scarcity Is Real, but Economics Still Decide
SanDisk reported $8.97B of revenue, an 84.6% adjusted gross margin, and a new $14B buyback. CoreWeave’s priority-access agreement with Solidigm further indicates that storage has joined compute as an AI bottleneck.
Even here, scarcity did not eliminate execution risk. SanDisk’s Q1’27 revenue guidance of $10.3B to $10.8B missed the $11.1B estimate.
There is credible evidence that investment is already creating value. Atlassian reported $1,766 million of revenue, up 28%, with cloud revenue up 31% and remaining performance obligations up 44%. Twilio generated $352.6M of free cash flow while raising FY26 guidance. Shopify produced $654M of free cash flow on $3.58B of revenue, up 34%. Airbnb beat the cited revenue, EPS, and gross booking value estimates.
Those are operating results, not merely an AI narrative. But strong demand and strategic importance still do not establish attractive returns on committed capital.
In the next phase of the AI boom, being right about demand is merely the entry ticket. Returns will depend on financing, asset life, customer concentration, and the price paid.