Hi, I'm Forrest

Tag: tulips

  • The Alchemy of Exuberance (Part 2)

    Structural Fragilities, Accounting Distortions, and the Systemic Risks of the Artificial Intelligence Cycle – Part 2

    Executive Summary: The “Rhyme” Has Become the Rhythm

    The financial architecture surrounding the AI sector has moved beyond simple exuberance into a phase of structural codependency. While the underlying technology remains transformative, the capital structures supporting it have effectively “inverted” the standard relationship between customer and vendor. We are now witnessing a market where the largest customers of AI infrastructure are funded directly by the vendors selling it, creating a closed-loop revenue system reminiscent of the telecom “capacity swaps” of 2000.

    This report updates the previous analysis with data from late 2025, specifically addressing the $45 billion capital-revenue swap between Microsoft, NVIDIA, and Anthropic; the “Pseudo-Acquisition” regulatory arbitrage; and the fracturing of the SaaS business model.

    I. The Circular Revenue Economy: The “Inverted” Deal Structure

    The most significant development in Q4 2025 is the industrialization of the “Round-Trip” transaction. The distinction between “Strategic Investment” and “Revenue Procurement” has dissolved.

    The $45 Billion “Inversion”: Microsoft, NVIDIA, and Anthropic

    On November 18, 2025, the market witnessed the apex of circular financing. Microsoft and NVIDIA announced a combined $15 billion investment into Anthropic. Simultaneously, Anthropic committed to purchasing $30 billion of computing capacity from Microsoft Azure (powered by NVIDIA GPUs) over the next 8+ years.

    The Mechanism:

    This deal “inverts” the typical capital cycle. Instead of a customer raising capital to buy a product, the vendor provides the capital to creating the booking.

    1. Cash Out: Microsoft/NVIDIA transfer $15B cash to Anthropic (Asset: “Equity Investment”).
    2. Contract Signed: Anthropic signs a $30B cloud commitment.
    3. Cash In: Anthropic pays Microsoft/NVIDIA back over time for compute services (Revenue: “Cloud/Hardware Revenue”).

    The “Rhyme”:

    This mirrors the vendor financing of the late 1990s telecom boom, where equipment makers like Lucent lent billions to startups to buy their own switches. The risk is that Microsoft and NVIDIA are effectively capitalizing their future revenue on their own balance sheets today. If Anthropic fails to commercialize its models profitably, the “revenue” Microsoft booked is simply a return of its own investment capital, disguised as growth.

    The Fragility of “Anchor Tenants”

    This structure suggests that organic demand from non-subsidized enterprises may be softer than headline numbers indicate. If the largest growth drivers in Azure and Data Center revenue are companies whose bills are paid with the cloud provider’s own equity checks, the “real” market clearing price for AI compute is opaque.

    II. The “Pseudo-Acquisition”: Regulatory Arbitrage via Acqui-Hires

    A new form of financial and legal engineering emerged in 2025 to bypass antitrust scrutiny: the “Reverse Acqui-hire.”

    The Pattern:

    Large Tech companies (Microsoft, Amazon) want to acquire AI startups (Inflection AI, Adept AI) but cannot pass FTC/DOJ review.

    1. The Licensing Fee: The acquirer pays a massive “licensing fee” (e.g., Microsoft paid Inflection $650 million) for access to models.
    2. The Mass Hire: The acquirer hires the CEO and the majority of the technical staff.
    3. The “Zombie” Shell: The startup remains technically independent (avoiding merger review) but is operationally gutted, using the licensing fee to pay back VC investors.

    The “Rhyme”:

    This rhymes with the “off-balance-sheet vehicles” (SIVs) of 2008. Just as banks moved assets off-books to avoid capital requirements, Big Tech is moving acquisitions “off-books” to avoid regulatory requirements. The FTC has already launched probes into these deals as of late 2025, arguing they are de facto mergers disguised as licensing deals.

    III. The GPU Asset Bubble: Shadow Inventory and Grey Market Crash

    The narrative of “infinite demand” for H100/H200 GPUs is colliding with the reality of product cycles.

    The Grey Market Signal

    Reports from late 2025 indicate a crash in the “grey market” price for NVIDIA H100s in China and other restricted regions. Prices for hoarded chips have dropped by ~10-20% as speculators rush to offload inventory before the H200 and Blackwell architectures achieve volume.

    The Collateral Problem

    This price erosion threatens the “Neocloud” financing model. Firms like CoreWeave and Lambda have raised billions in debt (e.g., CoreWeave’s $7.5 billion facility with Blackstone) collateralized by these chips. If the H100 is repriced from a “cash-flowing asset” to “obsolete inventory” faster than the loan amortization schedule, the collateral coverage ratios on these loans could breach covenants.

    This creates a “subprime hardware” risk profile in private credit markets.

    IV. The Depreciation Reality Check: Amazon’s 2025 Flip-Flop

    The “Asset Life Creep” noted in previous reports has hit a wall of technological reality. Companies can extend accounting lives on paper, but they cannot extend technical relevance in the rack.

    Amazon’s 2025 Adjustment:

    In a revealing move, Amazon increased the useful life of “heavy equipment” (power/shells) to 13 years (adding $0.9B to income), but simultaneously shortened the useful life of servers/networking gear from 6 years to 5 years (reducing income by $0.7B).

    The Implication:

    This admission that AI servers become obsolete faster than expected (5 years vs 6) directly contradicts the industry trend of extending lives to smooth earnings. It suggests a “Depreciation Cliff” is arriving. As the H100 generation is replaced by Blackwell, the accelerated depreciation of the older fleet will act as a massive headwind to GAAP earnings, which “Non-GAAP” adjustments will likely attempt to strip out.4

    V. The Fracturing of SaaS: RPO Divergence and “Agentic” Pricing

    The software business model is undergoing a painful transition that is obscuring growth metrics.

    Revenue vs. RPO Divergence

    Snowflake’s fiscal 2026 data shows a widening gap between Remaining Performance Obligations (RPO) growth (33%) and actual Product Revenue growth (28%). This “backlog bloating” suggests companies are signing longer, back-loaded contracts to maintain the narrative of high growth, even as immediate consumption slows. Net Revenue Retention (NRR) has stabilized but at lower levels (125-127%), indicating the “upsell engine” is sputtering.5

    The “Agentic Pricing” Confusion

    Companies like Salesforce are shifting from “per seat” pricing to “consumption” or “per conversation” models (e.g., Agentforce at $2/conversation or Flex Credits).

    • The Risk: This shifts revenue from predictable subscriptions (SaaS) to volatile utility billing.
    • The Rhyme: This rhymes with the 2010s pivot to “usage-based” billing in cloud, which caused massive volatility. However, with “AI Agents,” the definition of a “conversation” or “outcome” is subjective. This opens the door for “breakage” accounting—recognizing revenue on unused “Flex Credits”—to smooth quarterly misses.

    VI. Private Credit: The “Zombie” Covenant

    The private credit market, now the primary financier of tech LBOs and AI infrastructure, has degraded credit standards to historically risky levels.

    Uncapped Add-Backs:

    As of late 2025, 54% of private lenders now accept loan docs with uncapped EBITDA add-backs.6 This means a borrower can define “EBITDA” effectively however they choose, adding back “synergies” and “restructuring costs” without limit to stay compliant with leverage covenants.

    • The Result: A company with 6x real leverage can report 4x “Adjusted” leverage.
    • The Rhyme: This is “Covenant-Lite 2.0,” echoing the pre-2008 LBO boom where “Pro Forma” earnings replaced actual cash flow in credit underwriting.

    Conclusion: The Upstream Narrative

    The financial engineering observed in November 2025 confirms that the accounting is now “upstream” of the story. The Microsoft/Anthropic “inversion” is the clearest signal yet: when the market cannot support the necessary capex through organic revenue, the vendors must synthesize the revenue through circular investment.

    Watch List for 2026:

    1. The Unwind: Regulatory actions against the “Pseudo-Acquisitions” (Inflection/Adept).
    2. The Write-Down: Major cloud providers forced to impair “older” H100 clusters as Blackwell scales.
    3. The Covenant Breach: A “Neocloud” missing debt payments as GPU rental rates compress below debt service costs.