Nvidia capex

BofA's token-deflation thesis (token prices -50% since summer, Ramp business-AI spend fading since August) gets its demand-side test at Microsoft's late-Oct 2026 earnings: Azure growth vs the ~45% constant-currency guide, FY capex path vs $115.95B, and any capex language change. A miss or capex trim would be the first hard confirmation that the token squeeze is reaching a major seller of AI capacity — the equity-layer falsifier short of an outright capex cut (watched separately). Amazon reports Oct 29; Alphabet and Meta the same week.

CoreWeave is the largest neocloud and Nvidia's most exposed demand vehicle: Nvidia holds ~10% (second-largest holder), supplies its GPUs, and has agreed to purchase otherwise-unsold capacity through 2032. This week: stock -25%, junk yields >15%, 5-yr CDS intraday record 1,194.5bps implying >60% 5-yr default probability, >$40B debt vs ~$12.9B expected 2025 revenue, $104B backlog. Fires on: a default/distressed exchange/restructuring, an explicit Nvidia backstop or guarantee of CoreWeave debt, or sustained CDS normalization (back below ~600bps) signaling the credit market has stepped back from the brink.

TSMC is the foundry-layer gauge for the AI capex story (see [[tsmc-foundry-capacity-ai-demand]]). Q3 2026 (Oct 15): record net profit NT$763.06B +68.7% YoY, revenue $46.72B +50.9%, Q4 guide $52.3-54.3B (~55-61% YoY), FY26 raised to ~44%, capex 2026-28 "significantly higher than the past three years," Wei: "We did not see any sign of slowdown" — but "prudent in our business planning going into 2027." Each quarterly print tests whether the demand-gap language holds and whether the 2027 prudence dissolves or hardens. Also watch: quantification of the 2027 capex step-up, and whether TSMC's own US data-center ground-truth checks surface any customer slippage.

ASML is the upstream physical-evidence gauge for the AI capex story (see [[asml-euv-order-book-ai-demand]]). Q3 2026 (Oct 14): record €11.7B sales +55%, third 2026 guidance raise to €44-45B, "supply as currently planned for both 2027 and 2028 will not fully meet the demand of our customers," order interest out to 2029-2030. Each quarterly print tests whether the multi-year order book is extending (bullish for the capex story) or whether 2029-30 interest fails to convert to bookings (the upstream falsifier). Also watch the 2028 capacity-expansion decision beyond the planned +30%.

Microsoft unveiled Maia 300 on Oct 12, 2026, claiming 40% more inference tokens per megawatt than "leading GPU-based systems" (unnamed) and up to 7x the rack-scale token throughput of Maia 200; Reuters reports deployment plans that could top 1M units in 2027 with a multi-gigawatt footprint. Verify actual 2027 deployment scale, independent benchmark confirmation of the efficiency claim, and any disclosed GPU order reduction at Microsoft attributable to first-party silicon — the volume-side test of the substitution threat to Nvidia.

FT (Oct 6): SpaceX seeking ~$40B debt — ~$10B bank loans + ~$30B IG bonds, Apollo to lead, Pimco in talks, close expected 2027 — mostly to buy Nvidia chips for Colossus (Anthropic/Google tenants; Vera Rubin exclusive). SpaceX 5-yr CDS hit a record ~197.6bps (+16.7bp) on the report; total debt would double to ~$80B vs Q2 AI revenue of $2.6B and a $1.3B AI operating loss. Fires on: (a) confirmed pricing (spread/coupon vs the +175bp debut print and the 2056 tranche's +231bp), (b) downsizing/abandonment. The largest single Nvidia-chip debt financing in talks; its clearing spread is the IG-tier datum for the AI-credit curve.

Crypto Briefing (Oct 10): Broadcom's 5-year CDS hit a record 136bps by early October (record ~122bps in August, ~82bps late July), with each fresh financing announcement pushing spreads higher rather than reassuring lenders. Broadcom is the guarantor of the $42B Anthropic senior tranche and the assembler of the >$50B OpenAI Nexus package — its CDS is the market's price for the vendor-financing layer. Fires if the spread exceeds 136bps; sustained relief below the August print would signal the vendor-financing model is regaining credit traction.

9fin (Oct 8): Tract Capital's two 200MW Reno-area data halls leased by Nvidia still lack air permits for gas turbines and backup generators (confirmed with Nevada regulators); NV Energy sued Tract in August over 2GW buildout cost-shifting. The $3.8B (due 2031) and $4.6B (due 2031) junk bonds have fallen to 90.53 and 91.72. Fires on permit grant (bond-recovery test) or denial/slippage (first Nvidia-tenancy credit event, hardening the 90-91 cent template).

FT (Oct 10, via GuruFocus): Nvidia in early-stage talks to acquire or increase its investment in Reflection AI ($800M already invested; $25B March valuation); structures include outright acquisition, acqui-hire, bigger equity check, or enhanced chip-supply agreements. Parallel talks: investment in Perplexity at >$30B valuation. Fires on confirmation with terms/size, or abandonment. Deepens the circular-financing layer ([[off-balance-sheet-ai-infrastructure-financing-models]]) and extends the Poolside/Groq pattern to the model layer ([[nvidia-reflection-ai-acquisition-talks]]).

Gov. Abbott ordered a pause on new Texas data-center permits for a power/water audit (9fin, Oct 8); Galaxy Helios' $3.5B junk bonds dropped 3-4 points on the news and a wave of Texas projects has stalled. Texas is the largest AI data-center buildout market. Fires when: (a) the audit concludes and permitting resumes; (b) the pause hardens into lasting restrictions/moratorium; or (c) concrete project cancellations result. Outcome directly shapes the US pipeline conversion rate ([[ai-monetization-proofs-and-infrastructure-bottlenecks]]).

Brookfield and Naver are closing in on a roughly $9B loan from five South Korean banks to finance GPU purchases for their 200MW Sejong AI data center (Nvidia partner announcement; Naver plans expansion to 1GW) — per 9fin, Oct 8. Fires on closing with disclosed terms (the bank-financed GPU-purchase datum at the national-champion tier) or abandonment. Extends the debt-funded GPU-demand flywheel ([[off-balance-sheet-ai-infrastructure-financing-models]]) to a new lender geography.

Nebius is looking to issue $3-4B of junk bonds with Goldman Sachs as bookrunner (9fin, Oct 8) — the next test of the neocloud junk tier after Applied Digital and Novva postponed offerings to Q4 amid the spike in bond pricing. Fires on pricing/closing (yield vs the 8.875%-11%+ neocloud prints: Volta ~11% loan, Coatue-Fluidstack >11%) or postponement/pull.

After withdrawing its $5.5B ASX IPO (failed at A$11, ~$30B valuation track; valued $10.5B in August with Jane Street/Blackstone), the Nvidia-backed data-center builder is exploring a $2-3B private raise (Bloomberg, Oct 9-10). Fires on: (a) a closed private round with disclosed size and valuation — the post-scare private-market clearing price for AI-infra equity; or (b) failure to raise, which would strand the 912MW pipeline's funding model.

CNBC (Oct 8): OpenAI is in early-stage discussions with investors about a new funding round of ~$30B (figure could change, no term sheet finalized), driven by investor demand — after the $122B March round and the $50B run-rate disclosure that knocked 3-8% off the AI complex. Fires on: (a) a confirmed round at/above/below $30B with disclosed terms and valuation; (b) downsizing or abandonment. Post-scare round pricing is the private-market vote on the demand base's funding capacity — the input that keeps the debt-funded chip orders (Broadcom >$50B Nexus, SpaceX $40B) serviceable.

WSJ (Oct 7): Broadcom assembling a package exceeding $50B to finance custom AI chip production for OpenAI, with Apollo and Blackstone in early-stage talks; the "Nexus" program targets multiple gigawatts through 2029. This is vendor financing of the leading custom-silicon alternative to Nvidia's standard hardware ([[custom-silicon-and-datacenter-hardware-competition]]) — layered on top of the $102B Anthropic exposure (guarantor/lender/lessor/supplier, [[off-balance-sheet-ai-infrastructure-financing-models]]). Fires on: (a) confirmation and disclosed size/terms; (b) first tranche pricing; or (c) abandonment/downsizing. Also watch whether Broadcom's CDS (record 122bps in August) and its 21.5%-below-high equity discount deepen as contingent liabilities stack.

TechTimes (Oct 6) flags that TPU collateral is structurally less fungible than GPU collateral — TPUs "cannot be repurposed for general-purpose compute the way a GPU can," with "essentially one buyer class: frontier AI developers" — while existing credit frameworks were "developed largely around Nvidia GPU collateral." The Volta 11% print is unrated GPU-collateral paper; Broadcom's A-minus guarantee masks what ASIC collateral would clear at unguaranteed. As deals price (Volta commitments due Oct 14; the $42B Broadcom-guaranteed senior tranche; Amazon's $8B SPV), test whether GPU-backed paper clears systematically tighter than ASIC-backed paper at equal credit tiers. If yes, Nvidia's ecosystem holds a structural cost-of-capital advantage in the financing flywheel ([[off-balance-sheet-ai-infrastructure-financing-models]]); if no, the 11% print is about AI risk generally, not collateral quality.

PGIM anchored the first CLO with an explicit cap on AI-related collateral (15% of pool, $500M deal, late Sept 2026); BofA estimates CLO chip/data-center exposure could reach $100B. Fires when a second or third CLO manager adopts an AI concentration cap, or CLO purchases of AI-linked loans visibly stall — the credit-channel rationing signal for the capex flywheel. Complements the existing hyperscaler SPV-spread watch.

Nvidia is reportedly considering investing up to $10 billion as an anchor investor in Anthropic's IPO (Reuters, Sept 11; reiterated in Oct 3 IPO coverage). Fires when Nvidia confirms, sizes, or explicitly declines the anchor stake. A confirmed $10B check deepens the circular-financing layer (chipmaker anchoring its own demand base's valuation, [[off-balance-sheet-ai-infrastructure-financing-models]]); a decline would be a notable public signal about how Nvidia values the durability of the demand base ([[anthropic-ipo-compute-obligations]]). IPO expected November 2026, after the US midterms.

Zitron's "Dead Money" (Sept 29) tallies ~64% of hyperscaler AI revenue coming from OpenAI+Anthropic and flags TickerTrends data showing Anthropic ARR plateauing since ~June with OpenAI showing "initial signs of a slowdown" — "the biggest sign is the most obvious one — the deceleration of Anthropic and OpenAI." Fires on confirmed deceleration: audited S-1/prospectus numbers showing sequential revenue decline, a credible ARR disclosure below the prior print, or a major compute-commitment renegotiation. The Anthropic S-1/IPO (expected ~Nov 15, watched separately) is the nearest hard-data event.

Reuters (Oct 1): three banking sources say Nvidia "may need to offer guarantees on all of its deals, or they may have to be backed by a revenue stream from investment-grade customers"; "tens of billions of dollars of loan deals in the pipeline are likely to have strong guarantees and contracts." Fires when the first $500B-platform deal discloses terms: expanded/full Nvidia guarantee or IG-revenue-stream backing = the flywheel only works with Nvidia absorbing residual risk (vendor-financing redux, circular-financing concern lands on NVDA's balance sheet); unchanged ≤25% residual guarantee = credit repricing of GPU collateral continues.

FT (Oct 1–2, 2026): Amazon in talks to move ~$8B of installed Grace Blackwell chips (12+ US data centers, five states) into an SPV funded by outside debt, lease the chips back, and offer investors up to 10% equity; talks ongoing, Amazon declined to comment. Fires when the deal is priced/closed (terms: spread vs investment-grade curve, tenor, residual-value assumptions) or explicitly abandoned. A closing extends the GPU-collateral flywheel to hyperscaler scale; an abandonment signals credit investors won't price GPU residual risk even against an AA credit.

Project Jupiter's 2.46-GW Bloom Energy fuel-cell microgrid awaits a New Mexico air-quality permit with a November deadline (Global Data Center Hub, Sept 26). Fires on the permit decision: granted = the power path clears ahead of the Feb 1, 2027 pipeline completion; denied/slipped = the force majeure hardens and the 89-91 cent loan print becomes the template for AI-infrastructure credit repricing.

FT (Sept 29): Nvidia in early talks with insurers (via broker Howden Re) on structures to insure neocloud loans against GPU-collateral shortfall, with possible syndication to hedge funds; depreciation data already supplied to at least one insurer. Fires when a first agreement is announced, a named insurer/structure/pricing is disclosed, or the talks are confirmed abandoned. A signed deal extends the financing flywheel ([[off-balance-sheet-ai-infrastructure-financing-models]]); a collapse would signal insurers won't price GPU residual risk.

The Information (Sept 27) reported China's MIIT told some Chinese companies it intends to approve purchases of Nvidia's RTX Pro 5500 workstation GPU by Alibaba, ByteDance and others, with executives expecting the chip to avoid US export restrictions. Fires when Beijing formally approves the purchases, orders are confirmed/placed, or the signal is walked back — the first test of whether a post-summit China channel reopens for Nvidia at the margin.

Huawei announced (Sept 17) it will launch the Ascend 960DT in Q1 2027 and Ascend 960PR in Q3 2027, claiming its "Tau Scaling Law" enables high-performing chips without Western lithography and ~100,000-processor "supernodes." First real test of whether China's domestic AI stack erodes Nvidia's position — inside China (already ceded under export controls) and in third markets. DeepSeek is working closely with Huawei on deployment.

Nebius raised Nvidia GPU pay-as-you-go rental prices 17–21% effective Oct 1 (second hike in three months) and CoreWeave is signing contracts at higher prices — the cleanest real-time gauge of AI compute demand. Fires if either neocloud CUTS rental prices, discloses slowing utilization, or misses on renewal pricing — the inverse tripwire to the hyperscaler capex-cut watch.

Tripwire from Zitron's Sept 22, 2026 "Where're All The AI Chips?" audit, which estimates $50–100B of GPUs warehoused at Microsoft alone and ~$234B complex-wide. Fires if any hyperscaler or neocloud discloses a GPU inventory write-down, materially discloses the share of purchased GPUs not yet deployed, or reverses a server useful-life extension (depreciation catch-up). Would convert the deployment-gap narrative into a fundamental earnings/credit event.

Fed Chair Warsh disclosed (Fortune, Sept 17) that the Fed has set up an internal task force on AI, due to report by the end of 2026, studying "the implications for our future policy conjuncture." A report that treats AI-buildout financing as a financial-stability issue (circular financing, SPV structures, hyperscaler leverage) could harden into supervisory or margin-policy pressure on the debt-funded portion of the capex story.

At Dreamforce, Nice's CTO said models "even one generation behind" are "highly performant and effective... everything beyond this point is icing on the cake," and enterprises described model-routing to the cheapest capable model. If broadly true, frontier-model churn may not translate one-for-one into frontier-chip pull — a potential soft ceiling on the capex story that is distinct from the safety-pacing debate. Worth testing against: hyperscaler inference-vs-training utilization data, token-growth disclosures, and whether Nvidia's "twice as many chips in 2027" claim survives contact with enterprise purchasing patterns.

The September 2026 FOMC hiked 25bp to 3.75%–4.00% and signaled one more increase this year (median projection 4.1%). The next FOMC decision (late Oct or mid-Dec 2026) is the second-order test for the AI trade: a second hike pushes the 10-yr (now 5.00%, highest since 2007) further up, compressing AI/growth multiples and raising the cost of the debt-funded portion of hyperscaler capex. A hold would relieve the immediate macro overhang.

The falsifier for the capex story (per Seeking Alpha's checklist): any hyperscaler (MSFT/GOOGL/AMZN/META/ORCL) guiding capex down, cancelling a data center construction deal or GPU order, or a delayed GPU cluster delivery. Until then, slowdown talk is just talk — Carlson: "I need to see something concrete… versus just talk."

Monitor the regulatory clearance (FTC/DOJ) and official closing of Nvidia's agreed $12.9 billion acquisition of Hugging Face.

Track whether Digital Realty and Nvidia successfully complete and launch the 96 MW power-flexible Aurora AI Factory in Manassas, Virginia, utilizing Emerald AI's orchestration software.

Monitor whether Nvidia officially cancels, revamps, or resumes its $36 billion NeoCloud revenue-sharing and credit-support program, or folds it into the $500 billion Wall Street platform.

Verify if Nvidia's full-year FY2028 gross margin settles in the guided 72%-73% range as server price hikes and Samsung HBM4E qualifications take effect.

Monitor when Nvidia officially qualifies and begins high-volume commercial shipments of Samsung's custom 8-layer NVHBM (HBM4E) chip, expected to ease memory supply bottlenecks.

Monitor Nvidia's Q3 FY2027 earnings release, specifically verifying if revenue hits the guided $108 billion and if gross margins land in the guided 74% range.

Monitor the official pricing, capital raised, and trading debut of SB Energy's IPO, which is heavily backed by SoftBank, OpenAI, and Nvidia's $105B guarantee.

Verify if Nvidia's late 2027 flagship GPU, the Vera Rubin Ultra, actually ships with the downgraded 192GB of HBM4 8-hi memory configuration as reported, or if memory supply improvements allow an upgrade to higher capacity or HBM4E.

Monitor if Nvidia's 5-year credit-default swap (CDS) spread crosses 100 basis points (or $100,000 annually per $10 million insured), indicating severe credit market distress or escalating circular financing fears.

Monitor the successful deployment of AWS's committed 2 million additional Nvidia GPUs (Blackwell Ultra, Rubin, and Rubin Ultra) across its global cloud infrastructure by the end of 2028.

Verify if Nvidia's full-year FY2028 revenue meets or exceeds the guided 70% growth rate as supply constraints are resolved.

Verify if Nvidia's gross margins bottom out in the guided 71%-72% range in Q4 FY2027, or if memory price pressures push margins lower.

Monitor whether the Green Chile Project pipeline is successfully completed by February 1, 2027, to power Oracle's New Mexico data center campus, or if further permitting setbacks delay Project Jupiter's online date.

Monitor the completion, power delivery, and first operational workloads of OpenAI's 4.25 GW Phase 1 data center at the PORTS-Pike Technology Campus in Ohio, backed by Nvidia's $105 billion guarantee.

Monitor when Lancium achieves its grid hookup milestones to unlock Nvidia's additional $1 billion contingent investment, and track the first deployment of GB200 or Vera Rubin units at the 1.2 GW Stargate 1 campus in Abilene, Texas.

Verify if the >15% server price hikes are successfully implemented starting early 2027 and whether they protect or expand Nvidia's gross margin above 74% in subsequent quarters.

Monitor whether hyperscaler bond spreads or SPV borrowing spreads widen by more than 25 basis points relative to the broad investment-grade corporate market in any single quarter, signaling severe credit market repricing of AI debt.

Monitor aggregate free cash flow across the eight largest global hyperscalers, tracking BofA's projections of negative $64B in 2026, negative $144B in 2027, and negative $186B in 2028.

Monitor the official launch, terms, and capital deployment of the first compute financing transaction under Nvidia's $500 billion partnership with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR.

Monitor Alphabet's quarterly SEC filings (Form 10-Q/10-K) starting in Q3 2026 to track the volume, pricing, and dilution of Class A and Class C stock sold under its $40 billion ATM program to fund its massive AI capex.

Monitor the completion and official launch of SK Telecom's first facility of its planned 2-gigawatt AI data center in South Korea, powered by Nvidia's Vera Rubin platform and SK Hynix's HBM4.

Monitor SpaceX's launch of Starmind satellites carrying Nvidia's Vera Rubin NVL72 rackscale solutions into orbit, scheduled to begin in 2027.

Monitor HBM (High Bandwidth Memory) component pricing, manufacturing yields, and supply bottlenecks (SK Hynix, Micron, Samsung) as a key driver of hyperscaler hardware procurement costs and potential margin pressure for Nvidia and its customers.

Monitor Meta's progress on building and commercializing its own cloud infrastructure to sell or lease excess AI compute capacity to outside customers, following Mark Zuckerberg's hints during the Q2 2026 earnings call.

Track Google Cloud's Q3 2026 operating margin to verify if the bridging strategy of using third-party capacity puts downward pressure on the 35.6% operating margin achieved in Q2 2026.

Monitor the official commencement and initial workload performance of Google's leased 110,000-GPU cluster at SpaceXAI's Colossus 2 data center, scheduled to begin in October 2026.

Monitor Oracle's deployment of its planned 50,000-GPU Helios supercluster powered by AMD Instinct MI450/MI455X GPUs, which is expected to begin in Q3 2026.

Monitor AMD's mass production manufacturing ramp and commercial availability of its Instinct MI455X accelerator and Helios rack systems, currently projected for Q2 2027.

Verify if Nvidia begins commercial shipments of the Kyber NVL144 rack-scale systems by late 2027 or if SemiAnalysis's predicted delay to 2028 is borne out.

Monitor the official commercial release and deployment of Qualcomm's 250+ core Dragonfly C1000 CPU in Meta's next-generation server fleet.

Monitor Qualcomm's progress toward commercial sampling of its High Bandwidth Compute (HBC) Gen 1 technology with the AI250 accelerator, which aims to bypass HBM memory bottlenecks.

Monitor the initial deployment of OpenAI's custom Jalapeño inference chip in gigawatt-scale data centers with Microsoft and other partners by the end of 2026.

Track Oracle's progress on its announced plan to raise $40 billion in debt and equity during fiscal year 2027, including its $20 billion ATM program, to fund its AI data center expansion.

Monitor Anthropic's transition from its confidential June 2026 filing to a public S-1, official pricing, and stock market debut (with expected valuation of $965B+).

Monitor Meta's potential multi-billion dollar equity offering to fund its $125B–$145B AI capex budget, following Alphabet's $85B template.

Monitor the official commercial release and shipment of RTX Spark-powered laptops and compact desktops from hardware partners like ASUS, Dell, HP, Lenovo, Microsoft, and MSI.