Editor's Note

AI demand moved from an industry forecast to a contracting, pricing, and financing story. Compute capacity is increasingly being presented as an investable asset class, while memory, optics, and power are capturing more of the economics. The decisive question is no longer whether demand exists, but whether the capital funding it can earn adequate returns.

Compute Becomes a Financial Product

Nvidia’s memorandums of understanding with Apollo Global Management, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs, and KKR contemplate mobilizing more than $500 billion for GPUs and data centers. Jensen Huang called technology chips an “investable asset class,” while Blackstone President Jon Gray compared compute financing with mortgages.

The comparison is appealing because demand is increasingly contracted. AWS reportedly has much of its capacity committed through the end of 2027 and much of 2028, often through five-year customer commitments. Microsoft reported commercial remaining performance obligations of $678 billion, while CoreWeave’s backlog reached approximately $104 billion before more than $25 billion of early-Q3 commitments.

But a GPU cluster is not a toll road. Hardware faces rapid technical replacement, and financing can manufacture purchasing capacity without manufacturing end-customer profitability. The encouraging evidence is that economic obsolescence may be slower than technical obsolescence. CoreWeave reportedly signed an A100 contract extending through 2029, even though NVIDIA introduced the GPU in 2020. Its average debt cost fell about 300 basis points year over year, reportedly saving roughly $1.1 billion annually.

Nebius supplied even stronger scarcity evidence. Q2 revenue reached $582.3 million, up 454% year over year, while annualized recurring revenue reached $3.0 billion. Around 70% of Q2 deals included customer prepayments, and management expects more than $9 billion of prepayments in 2026 against over $40 billion of commitments. Short-duration agreements reportedly command $40 million to $50 million per megawatt, versus $20 million to $25 million for agreements lasting 1 to 3 years.

Management said, “We could sell our entire 2027 capacity on these terms today. We are deliberately not doing so because we see higher value in retaining some capacity for immediate customer needs.” That suggests power available now is worth materially more than promised capacity later. It also means the asset is not merely the GPU. The advantage is the ability to secure power, complete construction, finance equipment, and deliver usable capacity on schedule.

Firebird’s plan to bring 250 MW of NVIDIA AI infrastructure to Armenia and Kazakhstan extends the same logic to national capacity. SpaceX is reportedly targeting 10 GW of AI capacity by the end of 2027. AI infrastructure is spreading geographically and institutionally, but not every announced megawatt has equal economic value.

Memory, Optics, and Power Take Their Share

The week’s more interesting shift was value capture moving beyond accelerators. Micron reportedly cannot meet even half of customer demand for data center memory. Its Strategic Customer Agreements cover three to five years, include binding annual “Take or PAY” commitments, and provide substantial upfront cash.

Goldman Sachs estimates memory will represent approximately 62% of the bill of materials for Nvidia’s Vera Rubin superchip. Morgan Stanley analyst Howard Kao estimates memory cost inside a Vera Rubin rack will increase 435%, versus 57% for the GPU. SK Hynix, which held 58% of the HBM market in the first quarter, is investing $720 billion in a network of memory factories. Micron is spending $50 billion on two Boise fabs and building a $100 billion New York campus.

Sandisk offers the most aggressive version of the contractual thesis. It has reportedly signed eight agreements representing approximately $94 billion in total contract value at floor pricing, with weighted-average duration exceeding four years and $16.5 billion in financial guarantees. JPMorgan estimates these agreements will cover more than 50% of FY27 bits and approximately two-thirds in FY28.

Long-term agreements, customization, and prepayments could make memory less commodity-like. They cannot abolish cyclicality if $720 billion and other major capacity programs eventually outrun demand. Contract floors protect pricing, but only while counterparties remain able and willing to honor them.

Optics currently shows similar scarcity. Lumentum’s fiscal Q4 revenue rose 109.3% year over year to $1.006 billion, while non-GAAP gross margin expanded from 37.8% to 50.4%. CEO Michael Hurlston said, “we are way behind our shipments on high-powered lasers.” Coherent CEO Jim Anderson described “exceptional customer demand,” and Applied Optoelectronics reported an order book above $200M. Here too, the opportunity is paired with an execution test because AAOI plans to raise capacity from approximately 200,000 units per month to more than 650,000 by the end of 2026.

Power may be the deepest bottleneck. Nebius switched its completed Vineland facility to Bloom Energy fuel cells, while SpaceX reportedly is relying more heavily on natural gas. That suggests grid access and on-site generation are becoming part of compute economics rather than background utilities.

Scale With and Without Scarcity

Netflix provided a useful contrast to the infrastructure frenzy. Since 2021, its cash content spending has grown at a 2% annual rate while EBIT margins rose from 21% to approximately 31.5%. The company converts approximately 90% of earnings into free cash flow, primarily used for repurchases. Bill Ackman returned after an approximately 50% decline from Netflix’s June 2025 high of $134 reduced the forward earnings multiple from more than 40 times to 21 times.

That is scale converting into cash rather than scale demanding ever more capital. AI infrastructure may ultimately achieve similar operating leverage, but current evidence is dominated by backlog, contracted power, prepayments, and financing. Netflix’s economics are already visible in margins and free cash flow.

Software also pushed back against the idea that models will capture all the value. Microsoft said customers building with models from multiple providers increased 5x since the start of the year. Uber’s partnership with Pony.ai to deploy more than 2,000 robotaxis across Europe supports the possibility that workflow ownership and distribution can aggregate competing technology suppliers. The infrastructure layer looks scarce today, but customer access may prove more durable once capacity becomes plentiful.

Counter-Thesis and Risk Watch

The strongest counter-thesis concerns systemic financing, not absent demand. George Gammon argued that Nvidia and BlackRock’s $500 billion memorandum could spread AI credit risk through securitized loans, insurers, pension funds, and banks. Separately, @value-investing cited JPMorgan’s estimate that $4.1 trillion of $5.5 trillion in AI capital expenditure will be debt-financed. If borrowers remain unprofitable and collateral depreciates quickly, the “investable asset class” framing becomes the mechanism through which technology risk enters the financial system.

The macro evidence adds fragility without establishing a collapse. ClearValue Tax reported that July lost 23,000 jobs versus expectations for 83,000 added, while May and June were revised down by 103,000 combined. Technology recorded 149,023 job cuts through July 2026, with AI accounting for 33% of July cuts. The same channel’s claims that inflation is “totally fabricated” and closer to money-supply growth are opinions, not established facts.

Geopolitical Economy Report argued that China’s open-source progress could weaken the economics of leading US platforms, while also describing escalating conflict with Iran and gradual de-dollarization. Its cited claims require independent substantiation, but the competitive argument is material: cheaper models can expand token consumption while reducing the pricing power needed to support the infrastructure beneath it.

SemiAnalysis reportedly cut its 2026 Broadcom CoWoS forecast from 250k to 215k wafers and TPU v7x Ironwood production from 3.2mn to 2.7mn units. A disputed interpretation is that Google may be shifting suppliers; an alternative is a temporary CoWoS-S constraint. For Broadcom, AMD, and Google, distinguishing packaging scarcity from strategic share loss is now a consequential test.