The first production Vera Rubin systems have reached Microsoft data centers. Yet the announcements disclose neither their cost nor the return Microsoft expects to earn.
That gap between physical progress and financial evidence is now the central issue for AI infrastructure investors.
Delivery Is Proven, Attractive Economics Are Not
NVIDIA’s Vera Rubin platform is ramping into full production, and Microsoft CEO Satya Nadella confirmed that the first production systems had arrived at Microsoft data centers. NVIDIA and Microsoft have moved beyond planning and into deployment.
But delivery is not the same as attractive economics.
The announcements provide no rollout timing, customer availability, pricing, performance, power consumption, or expected Microsoft return on investment. They also do not establish when material revenue will be recognized or whether the resulting Azure economics will justify the investment.
Bigger Commitments Mean Bigger Unpriced Risks
Kalshi Finance reported that Broadcom’s latest AI financing package could reach nearly $100,000,000,000. No original source, structure, or counterparties were supplied, and it is unclear whether Broadcom would provide, receive, or arrange the financing.
A package approaching $100 billion could redistribute risk across the AI supply chain. Its value cannot be assessed without knowing what is being financed and whose balance sheet carries the exposure.
Capacity Models Are Not Execution Evidence
The trillion dollar case for Nebius assumes 7.5 gigawatts of active infrastructure by 2030, approximately $20 million to $22 million of revenue per megawatt, and a 6 to 8 times sales multiple. Those assumptions produce $150 billion to $165 billion in annual run rate revenue and a valuation between $900 billion and $1.3 trillion.
Nebius reportedly owns about 28% of ClickHouse and approximately 83% of Avride. Those holdings offer support, but they do not solve the larger execution problem.
Guided 2026 revenue is $3.0 billion to $3.4 billion, with a $7 billion to $9 billion annualized exit run rate. Reaching $150 billion would require roughly 20 times revenue growth over four years. Meanwhile, 3,400 megawatts of modeled capacity have not been allocated to named sites or contracts.
The model describes what fully utilized infrastructure could be worth. It does not establish that customers, financing, construction, and power will arrive on schedule.
Contract Value Is Not Shareholder Value
IREN’s more than $14 billion of secured contracted revenue looks striking beside its approximately $13 billion market capitalization. But revenue and equity value are not interchangeable.
The promoted $150+ share opportunity depends on contract duration, margins, counterparties, capital requirements, financing, dilution, and profitable execution. None of that information was supplied. The relevant question is how much capital IREN must invest to deliver the contracts and how much cash ultimately reaches shareholders.
Operating Progress Still Faces the Valuation Test
There is a credible counterargument. Early Vera Rubin deployment, IREN’s contracted revenue, Nebius’s planned capacity and investment holdings, and Micron’s reported earnings transformation could be the beginning of economics that eventually validate today’s ambitious valuations.
Micron’s reported results demonstrate extraordinary operating leverage. From two years ago to August 21, 2026, the share price reportedly rose from $109 to $946. EBITDA increased from $6 billion to $68 billion, while EPS moved from a loss of $1.40 to positive EPS of $45. The claimed forward P/E is 7.
However, the reporting periods and forecast methodology were not supplied. A low multiple on peak cycle earnings can still be expensive.
Micron’s CFO also said the company eventually expects to return 100% of excess cash to shareholders, principally through share repurchases after retaining appropriate liquidity. That is a meaningful capital allocation signal, but repurchases create value only when shares are bought below intrinsic value.
The Cost of Capital Is the Final Hurdle
Long-dated yields across the US, UK, France, and Japan are around 4% to 5%. The five-year Treasury offers 4.3%, and the two-year offers 4%.
At those yields, investors do not need heroic assumptions to earn a return. Every AI infrastructure project must compete against a meaningful hurdle rate.
AI infrastructure no longer needs to prove that machines can be delivered. The trillion-dollar question is whether the cash flows can outrun the cost of building and financing them.