Big Tech AI Strategy Brief · Nov 3–16, 2026

Models Got Cheaper and Wider, While the Money to Run Them Came From Markets

A fortnight with no US earnings: the major US and Chinese tech groups shipped models, borrowed to build and prepared one very large listing. Launches, partnerships, people and capex, compared side by side.

How to read this page. Every figure carries its basis (reported, estimate, upcoming, preview, GA, vendor-reported, independent, reported by press) and its date. Status is judged on the run date, Nov 17, 2026. The page describes; it is not investment advice.

60.0 vs 64.6Gemini 4 Carbon vs Claude Fable 5.2, independent index, Nov 12
$131B+announced loans, bonds and equity, third-party tally (plans, not proceeds)
$64.35Btop of Anthropic's filed IPO range; trading expected Nov 19 (upcoming)

Divrite Research · Status as of Nov 17, 2026

Cover visual: a skyline of abstract towers linked by soft arcs of light
Figure 1 · Cover of the Nov 3–16, 2026 brief
+94%Baidu AI Cloud Infra revenue to RMB 8.2B (reported, Nov 16)
$2 → $4Carbon input price per 1M tokens on Dec 1 (GA, Nov 10)
EUR 9BMeta first euro bonds, Nov 4; $15B of dollar bonds on Nov 3
$26Bseven suppliers' private placement in Anthropic's IPO (filed Nov 10)
~$3.2Toff-balance-sheet AI commitments (estimate, Nikkei)
−4.34%Nvidia shares on the day Huang resigned (reported, Nov 12)
Status labels: Reported / in effectAnnounced / preview / stagedIndependentPress report / vendor-reported / estimateUpcomingContext
01The verdict

Models got cheaper and wider while the money to run them came from markets

This window had no earnings from the US majors; the earnings-week figures are in the previous issue and appear here only as context. What happened was shipping, borrowing and one very large listing. Pick a takeaway to see its evidence.

Google opened its frontier model to paying customers, but an Anthropic model still tops the independent scale

Gemini 4 Carbon became generally available in the paid Gemini API on Nov 10 at $2 and $10 per million tokens, doubling to $4 and $20 on Dec 1; Gemini 4 Hydrogen followed on Nov 12 at $1.50 and $7.50. On Artificial Analysis Intelligence Index v5 (readings of Nov 12) Carbon scores 60.0 against 64.6 for Claude Fable 5.2 and 59.1 for Claude Opus 5.5, so no Big Tech in-house model leads. Hydrogen's per-token prices are 15 times those of Muse Ember.

Carbon GA · Nov 10Index 60.0 vs 64.6 · independent, Nov 12Hydrogen $1.50 / $7.50 · released Nov 1215x Ember's price
See the model comparison

The AI build moved from cash flow to the bond and equity markets

Meta priced $15 billion of dollar bonds on Nov 3 (annual coupon interest $990.1 million) and its first euro bonds, EUR 9 billion, on Nov 4 at about 1.4 times Alphabet's risk premium. Alphabet raised about $3.8 billion in its first Hong Kong and Singapore dollar bonds. A third-party tally puts announced loans, bonds and equity at more than $131 billion, and off-balance-sheet AI commitments at about $3.2 trillion (estimates).

$15B + EUR 9B · pricedAlphabet ~$3.8B · Nov 12$131B+ · third-party tally$3.2T · estimate
See the financing

Anthropic's IPO turns suppliers into shareholders

The amended S-1 sets 825 million shares at $70 to $78, up to $64.35 billion, and a $26 billion private placement by AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC; trading is expected on Nov 19 (upcoming). Amazon and Google together own more than a quarter of Anthropic, per the New York Times, and its compute commitments are about $895 billion to $901 billion.

825M shares · $70–$78 · filed Nov 10$26B placement · per the filingNasdaq trading · upcoming, Nov 19
See the IPO structure

Monetisation is becoming metered

Microsoft extended unpublished fair-use limits on Copilot Chat to unlicensed users (Nov 2 and Nov 9) with a $30-per-seat upgrade path. Amazon's Bedrock page lists a +75% Priority tier and a −50% Flex tier (secondary source). Meta's Ember scores 36 on the independent index at $0.14 per task against $0.41 for Haiku 5.5.

Copilot limits · in effect Nov 9$30 per seat · upgrade pathBedrock +75% / −50% · secondaryEmber $0.14 vs $0.41 · independent
See the company updates

The Chinese majors reported or shipped, and the chip story is still press-reported

Baidu's Nov 16 results show AI Cloud Infra revenue up 94% to RMB 8.2 billion while AI-native marketing fell 15%. Tencent's full Hy4 shipped on Nov 4 with weights still planned. DeepSeek's V5 on 160,000-plus Huawei Ascend chips is reported, not confirmed. Tencent (Nov 18) and Alibaba (Nov 19) report after the window.

Baidu · reported Nov 16Hy4 · shipped Nov 4V5 on Ascend · reported, unconfirmedTencent / Alibaba · upcoming
See China's majors
The window's pattern: models got cheaper and more widely available, while the money to run them came increasingly from markets, not cash flow.

What changed at a glance

CompanyChangeDateStatus
02US platforms, company by company

US platforms monetise by meter and ship at both ends of the price ladder

Microsoft and Amazon had no earnings in the window, so their news is about how each meters and sells AI capacity. Alphabet shipped a gated frontier model and a cheap everyday model in three days; Meta paid for and packaged its push; Apple went for distribution. Choose a company.

Dated timeline (2026)
  • Oct 26ContextVoluntary exit offer for about 60,000 eligible employees (roughly 27% of the reported workforce) in the US, Canada, UK and India; 3,000 or fewer expected to accept; decisions to early December (secondary reports).
  • Nov 2In effectCopilot Chat fair-use limits start for commercial users without a Microsoft 365 Copilot licence, in tenants of 2,000 or fewer seats.
  • Nov 3ReportedMessage Center items MC1536163 and MC1536158 published; The Register relays Microsoft's Partner Center announcement on Nov 5.
  • Nov 9In effectLimits reach larger tenants; caps are unpublished and admins cannot see or change them.
  • Nov 17–20UpcomingMicrosoft Ignite in San Francisco; Satya Nadella opens. Anything announced there is outside this issue.
  • Dec 1UpcomingReported delayUsage-based billing becomes the default for new CSP Copilot Business subscriptions, moved from Nov 2 (CRN via WindowsForum).
$30
Per user per month, annual billing
$31.50 billed monthly with an annual commitment · WindowsForum's reading of Microsoft pricing
$180,000
Per year for 500 upgrades
500 x $30 x 12 · the same source's arithmetic
Unpublished
Numeric fair-use caps
The limit notice usually carries an upgrade prompt
$25B
Bonds priced Oct 29
Context · first US investment-grade sale since 2017; expected to close Nov 4
Usage-based
Agent features such as Copilot Cowork
Billed separately from the per-user licence
3 tabs
Redesigned Copilot app
Home, Code and Autopilot, announced Sep 25 (context); Autopilot in private preview
Reading. The window's Microsoft story is monetisation, not models: fair-use caps that admins cannot see, a $30-per-seat upgrade and agents billed by usage. Microsoft sells access through seats and meters, so its lever is price discrimination on a captive base.
Dated timeline (2026)
  • Oct 23ContextAndy Jassy says 2027 AWS capacity is "largely sold out" (quoted by shattered.io).
  • Nov 3Secondary sourceBedrock page updated: Priority tier at a 75% premium to Standard, Flex tier at a 50% discount, Batch at half price for selected models.
  • Nov 9ReportedInfoQ reports TorchNeuron, a native PyTorch backend for Trainium introduced at PyTorch Conference North America 2026.
  • Nov 10FiledAnthropic's amended S-1 sets $70 to $78 per share; Amazon's stake is its largest AI exposure.
  • Nov 19UpcomingAnthropic's expected Nasdaq debut.
85%
Of AI developers write PyTorch
InfoQ via a secondary summary · the stated reason porting has braked non-Nvidia chips
+75% / −50%
Bedrock Priority / Flex
Treat as secondary until confirmed on the AWS pricing page
~$142.1B
Amazon's Anthropic stake at $74
TradingKey estimate at the midpoint; Alphabet about $132.4B
$250B+
If Anthropic reaches $1.4T
NYT: Amazon's $18B investment could exceed this; a different valuation basis, also an estimate
$110B+
Anthropic spend with Amazon
At least, over a decade · CNBC
up to $80B
Secured vendor-financing facility
Context · SEC filing before Nov 3; draws Q1 2027 to Q3 2028
Reading. TorchNeuron lowers the switching cost to Amazon's own silicon: existing PyTorch code runs on Trainium with virtually no changes, and early adopters include Cartesia and the Lieber Institute. Amazon sells capacity and chips, so its lever is utilisation of scarce compute, and its chip story depends on Anthropic staying a very large Trainium customer.
Dated timeline (2026)
  • Nov 9AnnouncedTPU 8tx (confidential computing) and TPU 8i Max unveiled; available next year (2027).
  • Nov 10GAStaged on Ultra surfacesGemini 4 Carbon open to paying developers and enterprises; Pixel Drop brings Gemini Intelligence to the Pixel 10 family.
  • Nov 11Rolling outGemini Live on Apple CarPlay for all Gemini users on iOS.
  • Nov 12ReleasedGemini 4 Hydrogen at $1.50 / $7.50. Same day, Reuters sees term sheets for Alphabet's first Hong Kong and Singapore dollar bonds, about $3.8B (reported).
  • Nov 12PreorderSamsung-built Gemini glasses: 41 g, 12MP camera, up to nine hours; in stores Dec 10 in the US and South Korea.
  • Dec 1UpcomingCarbon price doubles to $4 / $20.
$2 / $10
Carbon per 1M tokens to Nov 30
$4 / $20 from Dec 1; 95% off cached input · a 1M-input, 100K-output job goes from $3.00 to $6.00
$1.50 / $7.50
Hydrogen per 1M tokens
Per Shattered.io summarising Google; Google claims 2.5x the speed of Gemini 3.8 Flash
$395–489
Gemini glasses
Warby Parker from $395, Gentle Monster $449–489 · $395 undercuts Ray-Ban Meta Gen 3 ($449) by $54
~$12B → $135B+
TPU sales, 2026 to 2028
Bloomberg Intelligence estimate, excluding cloud rentals · 11.25x, about 235% a year
~3.5 GW
Anthropic TPU capacity from 2027
Google-Broadcom arrangement · nearly 100 TPU customers
~30%
Below the $408.61 52-week high
Reported only · shares rose over 2% premarket Nov 9; down 12.6% on Oct 29 (context). Not advice
Reading. Google's window is a pricing story at both ends. Sources disagree on Carbon availability: Shattered.io (updated Nov 10) calls it "invitation-stage" but describes only the Oct 28 announcement, while Google's posts and API release notes state GA for paying developers; this page follows Google. Forum posts on Nov 13–14 about "Unavailable" Gemini 4 models in Antigravity look like an entitlement problem, not regional gating. No dated report of TPU talks with Meta was found.
Dated timeline (2026)
  • Oct 26ContextEquity plan of Oct 26 (see the financing module); Q3 results of Oct 21 showed capex of $34.91B against operating cash flow of $32.79B.
  • Nov 3Priced$15B of US-dollar bonds in five tranches running 2031 to 2056; settlement Nov 5.
  • Nov 4PricedFirst euro bonds, EUR 9B in six tranches, record orders of EUR 53B.
  • Nov 6Rolling outSpecialised models behind Muse and Meta AI bring Active Memory and Active Focus to all users; Muse Wonder is not part of the rollout.
  • Nov 9AvailableMuse Ember, Meta's smallest and most efficient model, on the Meta Model API.
$990.1M
Annual coupon interest, dollar bonds
Sum of size x coupon across five tranches; weighted coupon about 6.60%
~1.4x
Alphabet's euro risk premium
Bloomberg's spread ratio; reoffer yields 1.07 to 1.14 times Alphabet's, 0.27 to 0.77 points higher
$0.10 / $0.50
Muse Ember per 1M tokens
Aimed at tool calls, coding subtasks, routing and extraction
36 at $0.14
Ember on AA Index v5, Nov 9
Independent · Haiku 5.5 scores 36.2 at $0.411; Ember cost is an estimate at proposed pricing
76% → 90%
Active Memory evaluation
Meta-reported, no independent evaluation · tasks with unexpected problems improved 67%
38% → 47%
Personal Superintelligence Index
Meta-internal composite of eight evaluations
Reading. Meta sold shares on Oct 26, dollar bonds on Nov 3 and euro bonds on Nov 4: three raises in nine days, priced above Alphabet's cost of debt. Meta reads Ember as matching Haiku's best score at about a third of the cost ($0.136 / $0.411 is about 0.33), which is consistent with the independent figures; the benchmark selection and framing are Meta's.
Dated timeline (2026)
  • Sep 14ContextSiri AI ships as a free beta with iOS 27 outside China, on Apple Foundation Models refined with Google Gemini.
  • Oct 31ConfirmedApple confirms China developer testing from Nov 4; Siri, Writing Tools and Visual Intelligence will hand some requests to Alibaba's Qwen.
  • Nov 4BetaFirst Apple Intelligence features for mainland China in the iOS 27.3 developer beta.
  • Nov 8Reported by pressBloomberg's Mark Gurman: Siri AI lead Mike Rockwell is seriously considering stepping down or leaving Apple in 2027.
  • Nov 9FiledShanghai's Cyberspace Administration publishes that Apple's generative-AI filing was amended to add iPad and Mac use cases; no launch date.
  • DecemberUpcomingPublic China release alongside iOS 27.3.
iPhone only
China build, for now
Excludes the new Siri AI, Image Playground and WWDC 2026 features · social-media summary
Qwen
Alibaba model inside Apple's China build
Confirmed by Apple on Oct 31
2 names
Reported successors
Amar Subramanya and Sebastien Marineau would reportedly share Rockwell's duties · reported by press
Accelerate
CEO John Ternus's reported push
Gurman: pressed software chief Craig Federighi to speed the AI roadmap beyond Siri · reported by press
Reading. Apple's window is about reach, not models. A filing is not a launch, and both people items are reported by press with no Apple comment found.
03China's majors

China's majors shipped and reported, while the chip story stays press-reported

Where the US groups financed and shipped, the Chinese groups split into three states: one that has reported (Baidu), two with announced dates (Tencent on Nov 18, Alibaba on Nov 19) and a chip story that is still press-reported (DeepSeek and Huawei).

Hy4's channels and the reported DeepSeek chip plan

Structure diagram of Tencent Hy4 channels and the reported DeepSeek V5 chip plan
Figure 2 · Confirmed channels (solid) and reported items (dashed)
Tencent Hy4, shipped Nov 4
The full version adds vision for the first time and can operate a computer for the user. It is live on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao and the WeChat Work assistant, with OpenRouter integration on launch day.
API price, in effect
RMB 6 per 1M input tokens, RMB 0.3 cached input and RMB 18 output, unchanged from the preview; 50% off through Dec 31, 2026.
"Open" is still a plan
Tencent says it plans to release the full model's weights; as of Nov 4 only Hy4-preview (770B parameters, 49B active, Apache 2.0, Aug 28) was downloadable.

Five groups, five states

Baidu Reported Nov 16

Q3 2026 results (quarter ended Sep 30): General Business revenue returned to 1% growth.

AI Cloud Infra RMB 8.2B, +94% · GPU Cloud +191% y/y, +43% q/q
AI-native marketing RMB 2.4B, −15% · AI Applications RMB 2.6B, about flat
Qianfan external token revenue +1,159% y/y · operating cash flow RMB 2.8B
Preview: ERNIE 5.5 Flash Preview, second among open-source models on Code Arena WebDev (vendor-reported); Token Exchange marketplace

AI Cloud Infra is 3.4 times AI-native marketing (8.2 / 2.4). A Tiger Securities preview expected near RMB 8B and +91% (estimate); Q2 was RMB 7.3B, +50% (context).

Tencent Hy4 shipped Nov 4Results Nov 18

Hy4 full version shipped with vision and computer use; weights planned, not yet released.

RMB 6 / 0.3 / 18 per 1M input, cached and output tokens; 50% off to Dec 31
AA Cyber Index 58 (Nov 13), second behind Kimi K3.1 Pro at 64 · independent
Capex: Q2 RMB 52.8B (+176%, Aug 12, context); Goldman estimates Q3 near RMB 66B (estimate)

Tencent's claim to rank among leading open-source models is its own positioning; no benchmark table placing Hy4 was found.

Alibaba Results Nov 19

Announced on Nov 6 that September-quarter results come on Thursday, Nov 19, before the US open, with a call at 7:00 a.m. Eastern.

No dated model or capex news inside the window in this material
Qwen is the model inside Apple's China build (confirmed Oct 31)
RMB 380B three-year cloud and AI plan dates from February 2025 (context)
Qwen4-Max Preview is subscription-only; Alibaba says it is retired after Nov 25 (secondary)

June-quarter figures are not reused as new.

ByteDance Gradual rollout

Seedream 5.5 image model released Nov 9 by the Seed team; Doubao said on Nov 10 its image design features were upgraded on top of it.

100 editing rounds with a stable picture and pixel-level preservation (ByteDance claim)
Native transparent (RGBA) output; better small text than Seedream 5.0 Pro
~$0.02 per image on a third-party API (WaveSpeed, via BytePlus listing)
No independent test and no head-to-head numbers yet; a Pro version is planned

DeepSeek / Huawei Reported, unconfirmed

Cailian Press and STAR Market Daily, citing supply-chain sources, reported on Nov 3 that DeepSeek plans to pre-train V5 mainly on Huawei Ascend 950 chips.

At least 160,000 Ascend 950DT at a new data centre in Ulanqab, Inner Mongolia (TrendForce, Dataconomy)
Operations from late 2026; V5 expected in 2027
~8 trillion parameters is an unverified figure (DigitalCitizen)
Pre-training has not started; neither company has confirmed
04Products and model comparison

On the independent scale no Big Tech model leads, and the cost spread is wide

Two independent snapshots from Artificial Analysis (AA) Intelligence Index v5 are used, each in its own view, because they are dated differently and the index moved between them. Index v5 launched on Oct 27 on a new scale, so scores do not compare with earlier versions. Vendor-reported results are kept apart.

4.6 points
Fable 5.2 ahead of Carbon (64.6 vs 60.0)
Fable 5.2 arrived on Nov 9 and ended the lead Carbon held on the Nov 3 and 4 readings (59 to 59.4).
~21x
Cost for the 9.6-point gap
GPT-6.1 Sol at 55 costs $1.27 per task; Fable at 64.6 costs $26.54 (26.54 / 1.27 = 20.9).
3 numbers
Carbon's cost per task is not stable
$5.48 on AA's German page (High), $6.59 for Medium on a comparison page, $7.76 via eesel. This page compares scores, not Carbon's cost.
ModelAccess and statusPrice (per 1M tokens)
Gemini 4 CarbonGA on API Nov 10; Ultra tiers staged$2 / $10 through Nov 30, then $4 / $20
Gemini 4 HydrogenReleased Nov 12$1.50 / $7.50, flat; 15x the lowest rates of Haiku 5.5, GPT-6 Luna and Ember
Claude Fable 5.2Released Nov 9 (gated Mythos 5.2 alongside)$8 / $40; fast mode $16 / $80
Muse EmberMeta Model API$0.10 / $0.50
Tencent Hy4Tencent Cloud and OpenRouter; weights plannedRMB 6 / RMB 18, 50% off to Dec 31
Qwen4-Max PreviewSubscription only (Token Plan); Alibaba says retired after Nov 25Per a secondary source
The 15x check: $1.50 / $0.10 = 15 and $7.50 / $0.50 = 15, matching VentureBeat's statement.

Google Carbon

Vendor-reported, Nov 10 #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model; #2 on the Multimodal Knowledge Agent benchmark at more than 6X lower cost than the top model.

Independent 60.0 on AA, second.

Google Hydrogen

Vendor-reported Ahead of Claude Sonnet 5.5 and GPT-6.1 Terra on seven of ten rows, but behind Sonnet 5.5 on Terminal-Bench 4.1 (57.0% vs 63.9%), OSWorld-2.1 and Terminal-Bench-Science.

Independent 51 to 54: 85% of Carbon's score at 24% of its cost per task (51/60 and $1.87/$7.76); Terra matches 51 at $0.96.

Tencent Hy4

Vendor-reported In Tencent's own tests Hy4 does not rank first on any of 12 benchmarks and sits 3rd to 4th of 7 compared models (explainx summary).

Independent AA index 35, mid-field; Cyber Index 58, second behind Kimi K3.1 Pro (64); leads CWE-Bench-AA at 76% pass@1 against 69% for DeepSeek.

Baidu ERNIE 5.5 Flash Preview and Meta Muse

Vendor-reported ERNIE is second among open-source models on Code Arena WebDev; a third-party write-up says it beats Opus 5.5 and GPT-6.1 Sol on some document and tool tasks but trails on complex reasoning. Meta's Active Memory and Personal Superintelligence Index results are Meta-internal.

Independent The only independent Meta numbers are the AA points: Ember 36, Wonder 44.

Build their own flagships

Google, Meta, Tencent and Baidu ship their own flagships; Google and Tencent also sell the model through cloud channels.

Partner, route or host

Microsoft's Copilot routes to OpenAI and Anthropic models rather than leading with its own; GitHub Copilot Fusion (Oct 28, context) orchestrates Anthropic and OpenAI models to balance cost and performance. Apple partners: Gemini-refined models in Siri, Qwen in China.

05Investments, partnerships and people

Anthropic's IPO turns suppliers into shareholders, and people and policy add risk

The financing pressure meets three concrete events: the terms of Anthropic's IPO, which makes three Big Tech groups shareholders and suppliers at once; a sudden leadership change at Nvidia; and a reported US programme that would name favoured AI companies.

The IPO and the supplier money loop

Structure of Anthropic's IPO and the supplier money loop
Figure 3 · Suppliers buy the stock and are paid by the commitments
Filed Nov 10 (Tuesday)
Amended S-1: 825 million Class A shares at $70 to $78, listing on Nasdaq as ANT. 801,351,354 are sold by the company and 23,648,646 by selling stockholders. Pricing expected Nov 18; first trading Thursday, Nov 19 (upcoming).
$26B private placement
AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC agreed to buy Class N stock at the IPO price, with no lock-up, per CNBC's reading of the filing. Part of the IPO money comes from the companies Anthropic pays.
Ownership
Amazon and Google together own more than a quarter (New York Times). Anthropic is committed to spend at least $110B with Amazon over a decade (CNBC).

Proceeds at the filed price

Drag the price to see 825 million shares at any point in the $70 to $78 range (our arithmetic).

$78
Price per share
$64.35B
Proceeds, 825M shares
$57.75B
At the bottom, $70
Reuters reports a valuation of up to about $1.4 trillion, which CNBC puts at $1.405 trillion fully diluted at the top of the range, against $965 billion in the May funding round. An earlier report on Nov 9 of about 867 million shares at $70 to $80 was described as inaccurate; follow the filed numbers.

Compute commitments and revenue

  • About $895B to $901B of compute commitments: $895.4B in one summary, $895B at Yahoo Finance, about $900.9B at shattered.io and TradingKey, "over $900 billion" at CNBC. The initial filing a week earlier said $417B. These are contractual commitments, not spend to date; the S-1/A is the authority.
  • $161B Broadcom of those commitments, against $63.9 billion of Broadcom's fiscal 2025 revenue (Yahoo Finance).
  • $33.242B revenue for the first nine months of 2026, up 1,206% (TradingKey; secondary, to be checked in the S-1/A).
  • Lower targets elsewhere: Bloomberg reported on Nov 9 that OpenAI and Anthropic were pursuing financings at lower-than-expected valuation targets after the market sell-off.

Nvidia: a CEO resigns Announced Nov 12

  • Announcement: Jensen Huang resigned as president, CEO and director, effective immediately; CFO Colette Kress is interim CEO while the board searches. Board chair Harvey Jones said operations, customer commitments and the roadmap continue without interruption (CNBC dates it to Thursday, Nov 12; an 8-K under Item 5.02 was filed).
  • Context: it followed days of reporting about Huang's personal relationships, including one with a businesswoman linked to a Chinese Nvidia customer; Huang apologised publicly on Nov 8.
  • Reported market and political effects: shares closed down 4.34%, erasing about $182B of value (after falling as much as 14.35%, about $600B, at the open); three House committees summoned Huang and the Senate Banking chair asked for testimony (CNN).
  • Next: results Nov 17 (upcoming), with Kress expected to run the call.

Policy: national champions Reported by press

  • Report, Nov 15: Reuters relays a Wall Street Journal report that the Trump administration plans to designate more than two dozen AI "national champions", including Nvidia, Amazon, Microsoft, OpenAI and Anthropic.
  • Benefits: expedited chip export-licence approvals, support for overseas operations and access to federal financing, per AI Weekly's summary.
  • Timing: Commerce Secretary Howard Lutnick is expected to unveil it within about a week (about Nov 22); the roster and criteria are not public. Google, Meta and Apple are not named in the reports reviewed.
  • Left out: social-media claims that the IMF flagged $1.4 trillion of circular AI deals could not be traced to an IMF document.
06Capex and financing

The AI build is now a financing story, and the stress shows at the edges

This window had no US earnings, so the capex actuals below are context from the previous issue. The window events are bond and loan filings, third-party tallies and Chinese pre-earnings estimates. Switch the chart to see each view and its status label.

Financing moves around the window

Built from one source (The Elec, citing Data Center Dynamics and SEC filings) except where marked; the status column separates the window from context.

CompanyInstrumentSizeDate / status
MetaUS-dollar bonds, five tranches$15BPriced Nov 3 (in window); the $97.5B equity plan of Oct 26 is context
MetaFirst euro bonds, six tranchesEUR 9BPriced Nov 4 (in window); orders EUR 53B (Bloomberg via Yahoo Finance)
AlphabetFirst HKD and SGD bonds, seven parts$3.8B (HKD 16B = $2B; SGD 2.28B = $1.8B)Term sheets seen by Reuters on Nov 12 (in window); orders above $7B
AlphabetEuro bond offering, 3.7% to 5.25%EUR 6BReported as being prepared; Bloomberg says Alphabet priced euro bonds on Nov 2 (just before the window)
MicrosoftFixed-rate bonds, 4.85% to 6.35%$25BPriced Oct 29 (context); expected to close Nov 4 (in window)
AmazonSecured vendor-financing loan facilityup to $80BSEC filing the week before Nov 3 (context); draws Q1 2027 to Q3 2028, floating rate, five-year maturity
Oracle$10B common stock plus $10B senior notes$20BDisclosed the same week (context)
The $131 billion headline (Data Center Dynamics) equals 80 + 25 + 20 + 6, so it adds the euro amount at face value. It is a third-party tally of plans, not of cash raised; it mixes loans, bonds and equity and names only Amazon, Alphabet, Microsoft and Oracle, so Meta's $15B dollar bond is outside it. The second wave of Meta's bond and Alphabet's first Hong Kong and Singapore dollar sales shows issuers diversifying currencies and investor bases.
~$200B
2027 gap left to fund
F.L. Putnam estimate: the big four spend $1.1 trillion in 2027, 45% more than 2026, against operating cash flow near $900B. Its "nearly $800 billion" for 2026 is rounded (1,100 / 1.45 is about $760B).
~$3.2T
Off-balance-sheet AI commitments
Nikkei, end of September for Alphabet, Meta, Microsoft, Amazon and Oracle: 90% above end-March (about $1.7T then). About 60% are long-term purchase agreements, 40% leases not yet commenced. Goldman's Amanda Lynam gives the same figure.
$296B → $440B
Hyperscaler debt supply, 2026 to 2027
Goldman Sachs projection (estimate). These obligations are disclosed in filing notes, not hidden, but they are not on the balance sheet.

The stress shows at the edges

  • CoreWeave: bonds originally issued at 9.67% now yield about 15% (The Elec, citing Fortune).
  • OpenAI: cut its fundraising target from $30B to $20B.
  • Goldman desk note (one view): after the US midterm vote, spending may be pulled forward into 2027 because states likely to tighten data-centre rules hold about 45% of planned US capacity.
  • Context from the Q3 reports: Microsoft's capex including finance leases was $45.2B and it lowered its 2026 outlook from $175B to about $170B; Alphabet spent $51.36B, held $195B to $205B and reported free cash flow of negative $9.01B against an expected negative $2.7B; Meta's Q3 capex was $34.91B against operating cash flow of $32.79B.

China: capex is rising, results are still ahead

  • Tencent, last reported (context): Q2 2026 capex of RMB 52.8B, up 176% year on year, reported Aug 12. Q3 results are scheduled for Nov 18 (upcoming).
  • Goldman estimate: Tencent's Q3 capex near RMB 66B (about $9.9B), to be checked on Nov 18. Short-sold Tencent shares rose from 3.44 million to 6.41 million between Nov 6 and 13, up 86%.
  • Alibaba: results follow on Nov 19 (upcoming); its RMB 380B three-year cloud and AI plan dates from February 2025 (context).
  • Baidu was the only Chinese major to report inside the window; its release gives no capex figure, so none is stated.
07Dated events

Dated events separate what is in effect from what is upcoming

Status is judged on the run date, Nov 17: dates before it are in effect or past; dates from it onward are upcoming. Filter by company and by status.

DateCompanyEventStatus
08Competitive outlook

Read capex as a financing story, keep models swappable and watch the dates

The window showed models becoming cheaper and more available, financing moving to the markets and several of the biggest dates falling just after it. Choose a reader to see what changed and what it implies.

What changed
  • Microsoft's unlicensed Copilot Chat limits are unpublished and admins cannot see them, with a $30-per-seat upgrade path.
  • Bedrock lists Priority at +75% and Flex at −50% (secondary source).
  • Carbon's price doubles on Dec 1.
Implication

Budget for metering, not just seats: plan the Copilot upgrade list before users hit caps, and lock Carbon usage plans before the Dec 1 step-up (a 1M-input, 100K-output job goes from $3.00 to $6.00).

$30 per seat · Microsoft+75% / −50% · secondary$2 / $10 → $4 / $20 · Google
What changed
  • Independent index: Fable 5.2 64.6, Carbon 60.0, Hydrogen 51 to 54, Terra 51; Ember 36 at $0.14 per task.
  • TorchNeuron lowers the cost of trying Trainium.
  • Hy4's weights are planned, not released.
Implication

Split workloads: a frontier model for hard jobs and a cheap tier (Terra at $0.96, Ember at $0.14 per task) for volume. Treat Hy4's open weights as planned, not available.

64.6 / 60.0 · independent$0.96 / $0.14 per taskHy4 weights · planned
What changed
  • $3.2 trillion of off-balance-sheet commitments (estimate) and a third-party $131 billion tally of announced raises.
  • Meta paying about 1.4 times Alphabet's euro spread.
  • Anthropic's suppliers are also its IPO buyers; Nvidia's CEO exit shows key-person risk.
Implication

The question is no longer capex size but who funds it and at what spread. Watch the circularity: supplier, investor and customer are the same firms.

$3.2T · estimate$131B+ · third-party tally~1.4x spread · Bloomberg

Company strategies side by side

CompanyBuild / partner / buyChipsFlagship in the windowFinancing stance
MicrosoftPartner: Copilot routes to OpenAI and Anthropic models; monetises through licences and usageMaia 300 announced Oct 12, production in 2027 (secondary)No in-house flagship in the window$25B bonds priced Oct 29 (context), closing Nov 4
AlphabetBuild, and sell TPU capacity to AnthropicTPU 8tx and 8i Max available 2027Carbon 60.0 and Hydrogen 51 to 54First HKD and SGD bonds, about $3.8B
MetaBuild (Muse)Not covered in the windowMuse Ember 36 at $0.14 per task; Wonder 44Equity plan (context), $15B dollar and EUR 9B euro bonds
AmazonPlatform plus chips; large Anthropic stakeTrainium with TorchNeuronNone in the windowSecured facility up to $80B disclosed before the window
ApplePartner: Gemini-refined models, Qwen in ChinaNot coveredNone in the windowNot covered; Siri leadership reports
BaiduBuild, plus a marketplace (Token Exchange)Kunlunxin M100 targeted for volume production by end-2026ERNIE 5.5 Flash PreviewCash flow positive: RMB 2.8B operating cash flow
TencentBuild and distribute (Hy4 in WeChat Work, WorkBuddy)Not coveredHy4: AA index 35, Cyber Index 58Results due Nov 18; Q3 capex estimate near RMB 66B
ByteDanceBuild (Seedream 5.5)Not coveredSeedream 5.5, no independent test yetNot covered

Three bets to track

Nov 18
Does Anthropic price inside its range?
And trade on Nov 19 (upcoming).
Nov 18 / 19
Do Tencent and Alibaba match the estimates?
Capex and cloud numbers against the estimates in the financing module.
~Nov 22
Does the champions list name Google, Meta or Apple?
The reports reviewed do not.
→
The question is no longer how big capex is, but who funds it and at what spread.This page describes the fortnight; it is not investment advice.