Editor's Note

AI demand became more tangible this week, expressed through long-term contracts, power commitments, custom silicon partnerships, and financing structures. The harder question is no longer whether infrastructure spending is real. It is whether the companies enabling that spending can earn attractive returns after accounting for financing risk, dilution, cyclicality, and a rising cost of capital.

From Silicon Scarcity to Balance Sheet Scarcity

The AI bottleneck is expanding beyond GPUs. Nvidia, OpenAI, and SB Energy plan an 8 GW compute campus in Ohio, with the first 4.25 GW expected to use Nvidia’s DSX platform and begin coming online in 2028. Supporting energy investment is expected to add at least 10 GW of new power. Nvidia reportedly invested $1.5B in SB Energy and agreed to credit support capped at $105 billion for the facility’s “land, power and shell.”

Amazon increased its planned Louisiana data center investment from $12B to $18 billion. Morgan Stanley projects US data centers will require around 68GW of power between 2026 and 2028, with a roughly 38GW shortfall after capacity already under construction, available, or contracted.

This changes the nature of the competitive advantage. Nvidia’s quarterly free cash flow reached $48.5 billion, 18 times three years earlier, giving it the financial capacity to help customers secure infrastructure. Broadcom is reportedly exploring a financing package that could approach $100B, including $60-70B of senior secured debt and approximately $30B of junior debt. Apollo and Blackstone are reportedly in talks, and Broadcom may guarantee part of the senior tranche.

These structures can accelerate deployment, but they also blur the line between supplying demand and financing it. The investment case now depends on who ultimately bears customer credit, utilization, and refinancing risk.

Demand indicators remain formidable. Broadcom reportedly has Q2 AI bookings above $30B, compared with $10.8B shipped, FY26 AI revenue guidance above $56B, and an FY27 target “in excess of $100B.” Anthropic’s reported revenue run rate increased from $9B at the end of 2025 to more than $65B in July 2026. Fabrinet reported Q4’26 revenue of $1.32B, up 45% YoY, while CoreWeave signed a multibillion-dollar AI cloud deal with Hudson River Trading.

The evidence supports a durable infrastructure cycle. It does not yet establish attractive returns for every participant funding it.

Google’s Supplier Strategy Reveals the New Custom Silicon Market

Marvell’s expanded Google relationship initially looked like a displacement threat to Broadcom. The details point instead to supplier diversification.

Google received the right to purchase up to 58,970,907 Marvell shares at $206.58 per share through August 2033. Only 1,360,867 shares vest on a fixed schedule during the first year. The remaining 57,610,040 shares are divided into 240 tranches, with one tranche vesting for every $500 million of Custom Products revenue Google generates for Marvell. Full vesting corresponds to a $120 billion revenue ladder, but it is not a revenue commitment.

The structure aligns dilution with commercial performance across AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. Broadcom reportedly remains Google’s primary partner for core TPU silicon under a long-term agreement announced in April and running through 2031.

BofA estimates Broadcom may retain 55% to 60% of TPU total addressable market value share, or approximately $250 billion to $350 billion, while Marvell captures 10% to 20%, or approximately $50 billion to $100 billion, and MediaTek captures 20% to 30%, or roughly $100 billion to $150 billion. Those are estimates, not contracted outcomes, but they show why Google can broaden its supplier ecosystem without declaring a single winner.

The meaningful signal is not Marvell’s nearly 10% or 13% rally, nor Broadcom’s nearly 5% or 6% decline. It is that hyperscalers increasingly want multiple suppliers across a widening custom silicon stack.

Memory Tests Whether This Cycle Is Structurally Different

Memory offers the cleanest test of duration. SanDisk reportedly has long-term agreements covering two-thirds of 2028 output and minimum contracted revenue of $93 billion. UBS estimates conventional DRAM gross margins, including at Micron, could reach an unprecedented 95% by 2027. J.P. Morgan reportedly forecasts approximately $1.8 trillion in combined DRAM and NAND revenue in 2028, including another 27% growth during that year.

Micron CEO Sanjay Mehrotra called memory “strategic infrastructure” and said AI is pulling “the entire memory hierarchy along with it, from HBM to DRAM to SSDs.” TSMC separately warned that “the industry is likely to face not only memory shortages but also tight ABF substrate supply over the next few years.”

Contracts and lead times suggest this cycle could last longer than investors using an old spot-price template expect. They do not abolish cyclicality. Projected margins of 80% or 95% invite capacity, substitution, and customer resistance. Micron’s more than $250 billion planned US manufacturing and R&D investment demonstrates how much capital must be committed before normalized returns become visible.

Retail positioning adds another warning. $DRAM assets reportedly reached a record $28 billion despite a 28% price decline, with $12 billion of inflows over two months. Since April 2nd, $DRAM reportedly attracted $27 billion and gained 98%. Strong fundamentals and crowded positioning can coexist.

Valuation Is Becoming a Cost-of-Capital Question

The 30-year Treasury yield reached 5.3% in one account, its highest level in over 19 years, while the 10-year moved above 4.7%. That matters because the AI buildout is becoming more capital intensive precisely when capital is becoming more expensive.

Nebius illustrates the tradeoff. Institutional ownership reportedly increased from 41.5% to 77.81% in under a year, while the company priced an upsized $5.0 billion convertible notes offering. Wolfe Research reportedly sees Nebius exiting 2030 with more than $41B of ARR from 5 GW of contracted power, but contracted power is not revenue. The economics depend on energizing capacity, maintaining utilization, controlling dilution, and preserving margins.

The same discipline applies outside infrastructure. Uber trades around 12 times projected 2027 free cash flow in one analysis, with more than 200 million monthly users and over 40 million daily trips. Another analysis reduced approximately $8 billion of annual free cash flow to roughly $4 billion after treating stock-based compensation as a real cost. Its moat depends on remaining the demand, distribution, and payments layer when autonomous fleets control the vehicles.

AI can transform an industry without making every exposed security attractive. As financing becomes part of the product, balance sheet strength and underwriting discipline become competitive advantages themselves.

Counter-Thesis and Risk Watch

The most relevant bearish argument came from the rates discussion. Bull Theory claimed $420 billion, and later $580 billion, had been erased from US stocks as yields rose, although the posts did not provide a supporting methodology. ClearValue Tax tied persistent pressure to roughly $2 trillion of annual borrowing and about $1.1 trillion of 2026 interest expense. George Gammon disagreed that Treasury buybacks determine long-term rates, arguing that expected growth and inflation ultimately dominate. Their mechanisms differ, but both keep the cost of capital at the center of the AI thesis.

Steve Eisman’s concern is more industry-specific: shallow LLM moats could produce a price war that travels into hyperscaler revenue and the estimated $400 billion of tech-related corporate bond issuance this year. The productivity evidence remains incomplete. Only 7% of companies reportedly have fully implemented AI, despite 56% having an AI account and about 30% reporting increased labor productivity.

@geopoliticaleconomyreport raised separate geopolitical risks involving alleged US interference in Latin America, the war against Iran, and potential depletion of Patriot interceptors. These are the channel’s claims, not independently established facts in the supplied material. The investable exposure is narrower: conflict, sanctions, critical minerals, defense access, and disrupted energy routes can quickly change financing and operating assumptions.

At the company level, @wallstreetmillennial reported that Rivian, excluding Volkswagen-related revenue and regulatory credits, remains at negative gross margins, loses about $10,000 per vehicle, burns cash at $3.5 billion per year, and has more than $1 billion of quarterly operating losses. That is the clearest reminder of the week: technological progress does not rescue a business whose unit economics and financing needs fail to converge.