TTN Research Alert: When beats don’t bid: what Vertiv, Oracle, repricing of AI impact on CRM and the “circular AI” flywheel say about the next leg; Markets seem to be starting to grade AI on FCF and physics, not poetry
Thursday, October 23, 2025 3:09:37 AMEST
1) The tape is whispering before the headlines do. Data center power-and-cooling bellwether Vertiv (VRT) has become a kind of Geiger counter for AI infrastructure demand. Its post-earnings reaction yesterday has been erratic even on strong print and enormous jump in organic orders, despite upbeat guidance—an early hint that perfection is priced in. Meanwhile Oracle (ORCL)—an unexpected AI market proxy thanks to its cloud/compute ambitions—slid after unveiling multiyear targets at its investor event, suggesting that even “AI winners” are being graded on cash flows, not capex poetry. The broader tape has started to punish AI-linked names on any whiff of profit-timing slippage, as we saw in early October when a single report questioning AI compute monetization knocked the sector. When strong results don’t reliably rally stocks, risk premia are changing. Reminder: in Aug, Altman himself said "yes, we are in a phase where investors as a whole are overexcited about AI" and that "someone is going to lose a phenomenal amount of money." All these happening amid various AI tech optimization announcements like Alibaba's Aegaeon computing pooling solution that it said led to an 82% cut in the number of Nvidia graphics processing units (GPUs) needed to serve its AI models and potentially reshaping AI workloads.
2) Beneath the surface: a spaghetti bowl of money flows. The current craze for “circularity”—equity, warrants, pre-buys and backstops that loop cash from supplier to customer and back—looks less like old-school vendor financing and more like a capital-intensive mutual-assurance pact. Flows among OpenAI, Nvidia, Microsoft, Oracle, AMD and CoreWeave map to a dense, reciprocal web: investment on one arrow, chip or capacity purchases on the return, sometimes with undisclosed counterparties. It’s not necessarily nefarious, but it makes final customer unclear and concentrates risk. If enthusiasm for data-center spend fades, suppliers could be hit twice—by lower sales and writedowns on stakes in their own customers. Circularity is virtuous on the way up—and vicious on the way down.
3) The right analogy isn’t crypto; it’s late-1980s Japan. There, keiretsu cross-holdings provided a cushion against market discipline—until mark-to-market rules and governance changes forced an unwind. Japan’s cross-shareholding ratio fell to ~7.6% by March 2004, a collapse of the old “we own each other” compact, puncturing valuation support mechanisms that had seemed permanent. The lesson: cross-ownership masks fragility until accounting and capital constraints compel clarity. Today’s AI “ecosystem” stakes rhyme with those old stabilizers.
4) Or think of Lucent—and the cost of selling to your own balance sheet. In the 1999–2001 telecom boom, Lucent extended billions in financing to customers (notably Winstar), boosting near-term shipments and “growth.” When those customers failed, Lucent’s losses ballooned (a $3.69 billion quarterly hit in April 2001), and the SEC later charged the company over revenue recognition tied to these practices. Financing your customers amplifies cycles. Today’s AI—via equity, commitments, take-or-pay lanes and capacity backstops—has echoes.
5) WorldCom’s “capacity” games supply the other cautionary tale. In the last bubble, carriers swapped indefeasible rights of use (IRUs) and tried to book non-cash “capacity swaps” as revenue; regulators and Congress eventually clamped down. The point isn’t that today’s AI players are doing that; it’s that revenue that depends on counterparties doing symmetric, financially circular things deserves a higher discount rate—especially when disclosure is thin and measurement slippery. Round-trip economics always look fine—until they don’t.
6) Enron teaches the accounting variant of the same lesson. Its “round-trip” trades inflated volumes without creating value, spurring a post-mortem regulatory architecture that still shapes energy and market-manipulation doctrine. Again, not an accusation—a reminder that complex, offsetting transactions can launder risk perception. In AI, the analog is pre-paying, equity-seeding, or backstopping the very demand you then report as momentum. If the economics are truly robust, circular props shouldn’t be necessary.
7) Meanwhile the physics are getting in the way. AI’s supply chain bottleneck is no longer just GPUs; it’s packaging and memory. CoWoS-class advanced packaging remains the choke point despite aggressive expansion, with industry trackers expecting capacity to roughly double in 2025 yet remain tight, while HBM4 timelines push meaningful relief into 2026. That’s supportive for pricing and for chip vendors’ P&Ls—but it caps the throughput of the entire build-out, extending the cash-burn window for downstream “AI application” stories that need cheap, abundant inference to earn margins. Bottlenecks keep hype hot—and delay cash conversion.
8) Then comes the grid. Data-center power demand is exploding, with S&P Global estimating +22% U.S. grid power for data centers by end-2025 and nearly 3× by 2030. Yet utilities face transformer shortages and 80–210-week lead times on large units, with DOE and NIAC flagging multi-year procurement delays. Interconnection queues are swollen—~2,300 GW of generation/storage seeking a hookup as of end-2024—and median queue-to-operation times creeping toward 5 years. This is the non-linearity investors keep underestimating: even if capital is infinite, electrons and copper aren’t.
9) Policy risk is growing, not receding. Washington keeps tightening AI-chip export controls, while Beijing has expanded critical-minerals restrictions (from gallium/germanium to broader rare-earths and strategic metals), directly threatening HBM, magnets, and advanced packaging inputs. Geopolitics can pinch the “AI stack” from both ends—chips out, minerals in—introducing fat-tail outcomes that valuation models rarely price correctly in bull phases. Cost of capital lives in the footnotes—but it trades in the tape.
10) “It’s all free anyway”—the software margin problem. At the application layer, the trend is toward bundling AI features at little or no incremental price into existing subscriptions (notably across Windows and Microsoft 365), while cloud-scale inference pricing faces competitive pressure. Free-to-cheap AI front ends push monetization to the back end—workflow redesign, vertical integration, or usage-based value shares—which takes longer and requires real switching costs. That’s why stock reactions to guidance that emphasizes capex first, cash later are getting harsher: investors have seen this movie.
11) Adoption is real but messy. Seven in ten enterprises say they’re using gen-AI in at least one function, per McKinsey, yet Gartner expects >40% of autonomous-agent projects to be canceled by 2027—a sober proxy for today’s integration friction, data curation costs, and controls. On the ground, making agentic systems talk to legacy, mutually suspicious systems can be slow and brittle, especially in regulated industries like law and finance. Hype compresses timelines; governance elongates them.
12) Why this matters for the “circularity” debate. When unit economics slip or power/talent/packaging constraints delay scale, ecosystem finance tends to grow faster than end-customer cash—because vendors must seed, subsidize, and pre-buy to keep the flywheel spinning. That’s exactly how telecom’s bubble metastasized from optimistic spreadsheets into balance-sheet contagion: the same dollar showed up as investment, then revenue, then collateral. The red flag isn’t coordination—it’s dependence. The day the music pauses, sequential arrows become circular risk.
13) What credible guardrails look like. Investors should demand segmentation that discloses third-party, arm’s-length revenue vs. ecosystem-supported sales, FCF coverage of capex (especially for data-center additions), signed utility interconnects with dated milestones, and term-priced compute commitments that survive price wars. Watch Vertiv as a power and thermal proxy; watch Oracle as a hyperscale-adjacent buyer/seller of compute; watch packaging and HBM lead-time prints for real capacity signals. If these indicators improve without circular scaffolding growing, the thesis is durable. If not, history’s rhyme gets louder.
14) The “bailout” temptation is real—but not for stockholders. Governments are moving to de-risk the grid (note the U.S. DOE’s reactor pilot program and **NuScale’s NRC design approval on May 29, 2025), and lawmakers are unlikely to permit a systemic power shortfall to kneecap digital competitiveness. But industrial policy backstops electrons, not equity valuations. If circular financing masks weak end-demand, policy won’t immunize P&Ls. The right question from the 1990s still applies: How does the technology support any business objectives, at a margin that clears capital costs?
- The bear case in one paragraph. Stocks fail to ramp on good news (VRT/ORCL signal), packaging and HBM stay tight until 2026, power and transformers lag, export controls and mineral restrictions bite, application margins compress via bundling, agentic AI stalls in integration, and circular cash flows obscure real demand. That cocktail doesn’t require fraud to reprice the AI complex; it merely requires time. Timelines are precisely what speculative cycles mis-price.
- The bull case in one paragraph. Supply bottlenecks keep prices firm for the core suppliers; software productivity gains compound invisibly and show up suddenly in 2026–2027; grid expansion accelerates (including advanced nuclear pilots) while export regimes stabilize; unit-cost curves for inference continue to fall; real-world agentic systems find durable, high-frequency use-cases. In that world, today’s circularity looks like **wartime finance—messy but necessary—**and investors are paid for enduring volatility.
- The median probable outcome? Likely not Armageddon, not nirvana—triage. Expect sharp dispersion: infrastructure names that can translate backlog into free cash flow despite grid delays; application players that prove durable price power; and tourists—firms whose AI narrative was a funding strategy. To separate them, follow the cash, power, packaging, and disclosure. This time isn’t different; the plumbing is.
~
2) Beneath the surface: a spaghetti bowl of money flows. The current craze for “circularity”—equity, warrants, pre-buys and backstops that loop cash from supplier to customer and back—looks less like old-school vendor financing and more like a capital-intensive mutual-assurance pact. Flows among OpenAI, Nvidia, Microsoft, Oracle, AMD and CoreWeave map to a dense, reciprocal web: investment on one arrow, chip or capacity purchases on the return, sometimes with undisclosed counterparties. It’s not necessarily nefarious, but it makes final customer unclear and concentrates risk. If enthusiasm for data-center spend fades, suppliers could be hit twice—by lower sales and writedowns on stakes in their own customers. Circularity is virtuous on the way up—and vicious on the way down.
3) The right analogy isn’t crypto; it’s late-1980s Japan. There, keiretsu cross-holdings provided a cushion against market discipline—until mark-to-market rules and governance changes forced an unwind. Japan’s cross-shareholding ratio fell to ~7.6% by March 2004, a collapse of the old “we own each other” compact, puncturing valuation support mechanisms that had seemed permanent. The lesson: cross-ownership masks fragility until accounting and capital constraints compel clarity. Today’s AI “ecosystem” stakes rhyme with those old stabilizers.
4) Or think of Lucent—and the cost of selling to your own balance sheet. In the 1999–2001 telecom boom, Lucent extended billions in financing to customers (notably Winstar), boosting near-term shipments and “growth.” When those customers failed, Lucent’s losses ballooned (a $3.69 billion quarterly hit in April 2001), and the SEC later charged the company over revenue recognition tied to these practices. Financing your customers amplifies cycles. Today’s AI—via equity, commitments, take-or-pay lanes and capacity backstops—has echoes.
5) WorldCom’s “capacity” games supply the other cautionary tale. In the last bubble, carriers swapped indefeasible rights of use (IRUs) and tried to book non-cash “capacity swaps” as revenue; regulators and Congress eventually clamped down. The point isn’t that today’s AI players are doing that; it’s that revenue that depends on counterparties doing symmetric, financially circular things deserves a higher discount rate—especially when disclosure is thin and measurement slippery. Round-trip economics always look fine—until they don’t.
6) Enron teaches the accounting variant of the same lesson. Its “round-trip” trades inflated volumes without creating value, spurring a post-mortem regulatory architecture that still shapes energy and market-manipulation doctrine. Again, not an accusation—a reminder that complex, offsetting transactions can launder risk perception. In AI, the analog is pre-paying, equity-seeding, or backstopping the very demand you then report as momentum. If the economics are truly robust, circular props shouldn’t be necessary.
7) Meanwhile the physics are getting in the way. AI’s supply chain bottleneck is no longer just GPUs; it’s packaging and memory. CoWoS-class advanced packaging remains the choke point despite aggressive expansion, with industry trackers expecting capacity to roughly double in 2025 yet remain tight, while HBM4 timelines push meaningful relief into 2026. That’s supportive for pricing and for chip vendors’ P&Ls—but it caps the throughput of the entire build-out, extending the cash-burn window for downstream “AI application” stories that need cheap, abundant inference to earn margins. Bottlenecks keep hype hot—and delay cash conversion.
8) Then comes the grid. Data-center power demand is exploding, with S&P Global estimating +22% U.S. grid power for data centers by end-2025 and nearly 3× by 2030. Yet utilities face transformer shortages and 80–210-week lead times on large units, with DOE and NIAC flagging multi-year procurement delays. Interconnection queues are swollen—~2,300 GW of generation/storage seeking a hookup as of end-2024—and median queue-to-operation times creeping toward 5 years. This is the non-linearity investors keep underestimating: even if capital is infinite, electrons and copper aren’t.
9) Policy risk is growing, not receding. Washington keeps tightening AI-chip export controls, while Beijing has expanded critical-minerals restrictions (from gallium/germanium to broader rare-earths and strategic metals), directly threatening HBM, magnets, and advanced packaging inputs. Geopolitics can pinch the “AI stack” from both ends—chips out, minerals in—introducing fat-tail outcomes that valuation models rarely price correctly in bull phases. Cost of capital lives in the footnotes—but it trades in the tape.
10) “It’s all free anyway”—the software margin problem. At the application layer, the trend is toward bundling AI features at little or no incremental price into existing subscriptions (notably across Windows and Microsoft 365), while cloud-scale inference pricing faces competitive pressure. Free-to-cheap AI front ends push monetization to the back end—workflow redesign, vertical integration, or usage-based value shares—which takes longer and requires real switching costs. That’s why stock reactions to guidance that emphasizes capex first, cash later are getting harsher: investors have seen this movie.
11) Adoption is real but messy. Seven in ten enterprises say they’re using gen-AI in at least one function, per McKinsey, yet Gartner expects >40% of autonomous-agent projects to be canceled by 2027—a sober proxy for today’s integration friction, data curation costs, and controls. On the ground, making agentic systems talk to legacy, mutually suspicious systems can be slow and brittle, especially in regulated industries like law and finance. Hype compresses timelines; governance elongates them.
12) Why this matters for the “circularity” debate. When unit economics slip or power/talent/packaging constraints delay scale, ecosystem finance tends to grow faster than end-customer cash—because vendors must seed, subsidize, and pre-buy to keep the flywheel spinning. That’s exactly how telecom’s bubble metastasized from optimistic spreadsheets into balance-sheet contagion: the same dollar showed up as investment, then revenue, then collateral. The red flag isn’t coordination—it’s dependence. The day the music pauses, sequential arrows become circular risk.
13) What credible guardrails look like. Investors should demand segmentation that discloses third-party, arm’s-length revenue vs. ecosystem-supported sales, FCF coverage of capex (especially for data-center additions), signed utility interconnects with dated milestones, and term-priced compute commitments that survive price wars. Watch Vertiv as a power and thermal proxy; watch Oracle as a hyperscale-adjacent buyer/seller of compute; watch packaging and HBM lead-time prints for real capacity signals. If these indicators improve without circular scaffolding growing, the thesis is durable. If not, history’s rhyme gets louder.
14) The “bailout” temptation is real—but not for stockholders. Governments are moving to de-risk the grid (note the U.S. DOE’s reactor pilot program and **NuScale’s NRC design approval on May 29, 2025), and lawmakers are unlikely to permit a systemic power shortfall to kneecap digital competitiveness. But industrial policy backstops electrons, not equity valuations. If circular financing masks weak end-demand, policy won’t immunize P&Ls. The right question from the 1990s still applies: How does the technology support any business objectives, at a margin that clears capital costs?
- The bear case in one paragraph. Stocks fail to ramp on good news (VRT/ORCL signal), packaging and HBM stay tight until 2026, power and transformers lag, export controls and mineral restrictions bite, application margins compress via bundling, agentic AI stalls in integration, and circular cash flows obscure real demand. That cocktail doesn’t require fraud to reprice the AI complex; it merely requires time. Timelines are precisely what speculative cycles mis-price.
- The bull case in one paragraph. Supply bottlenecks keep prices firm for the core suppliers; software productivity gains compound invisibly and show up suddenly in 2026–2027; grid expansion accelerates (including advanced nuclear pilots) while export regimes stabilize; unit-cost curves for inference continue to fall; real-world agentic systems find durable, high-frequency use-cases. In that world, today’s circularity looks like **wartime finance—messy but necessary—**and investors are paid for enduring volatility.
- The median probable outcome? Likely not Armageddon, not nirvana—triage. Expect sharp dispersion: infrastructure names that can translate backlog into free cash flow despite grid delays; application players that prove durable price power; and tourists—firms whose AI narrative was a funding strategy. To separate them, follow the cash, power, packaging, and disclosure. This time isn’t different; the plumbing is.