Most existing data center capacity is held by CSPs, while most new capacity in Texas is being added by neoclouds. CSPs can also add capacity outside the U.S. (see $GOOG recent 15bn investment in Malaysia)

Research|AIDC: Texas Permitting Risks Rise After Midterm Elections; 2027 Delivery Forecast Unchanged In the November 3 midterm elections, Democrats won control of both chambers of Congress. Texas was the biggest surprise: Talarico won the U.S. Senate seat, becoming the first Democrat to win the seat since 1988. Hinojosa defeated Abbott, who was seeking a fourth term, while several statewide offices, including lieutenant governor, attorney general and comptroller, also changed hands. Republicans are still expected to control the Texas Senate. The final composition of the Texas House and the comptroller race should be updated before publication based on the certified vote count. Data centers were one of the central issues in the election. Hinojosa pledged to impose a moratorium on new data centers until the legislature establishes safeguards; require projects to demonstrate that they will not raise residential electricity or water bills; prohibit developers from signing NDAs with local governments; and eliminate the sales tax exemption for data centers (6.25%, for up to 20 years, with an estimated two-year fiscal cost of approximately $3.2bn). The freeze was already underway. On August 3, Abbott halted the ERCOT large-load queue (approximately 474GW of applications, around 90% of which were data centers, or more than five times Texas’ historical peak load) and ordered a project-by-project review. ERCOT subsequently delayed Batch Zero in its new interconnection process. On September 21, Abbott also directed TCEQ not to issue any data-center-related permits until the ERCOT and TWDB audits are complete, with all state agencies suspending related approvals. The ERCOT audit is expected in early December, but its completion will not automatically restart permitting. Before the election, the market broadly expected Abbott to win re-election and was focused on when the freeze would be lifted after the audits were completed. Hinojosa’s victory could mean a longer freeze and tighter conditions for reopening the permitting process. However, she said during the campaign that she was not opposed to data centers. Her proposals focused mainly on protecting household utility bills and giving communities a say in site selection. The main impact of the election, therefore, is likely to be on how long the freeze remains in place and the conditions under which permitting resumes. Under our base case, we are temporarily maintaining our 2027 delivery forecast. Supporting that view will require a closer review of project permitting status: how much capacity has obtained key permits, how much Texas capacity remains subject to approval, and how much delay is already reflected in the original model. If the freeze lasts longer than assumed in the base case, or the probability of the downside scenario increases, expected deliveries could also be revised lower. New projects in Texas face greater exposure in 2028: most are likely to be delayed, while some demand may shift to other states. Detailed Report

Breaking down Anthropic’s 3Q26 inference costs, we can see that SpaceX captured the vast majority of the incremental spend, while AWS, GCP, and, to a lesser extent, Azure captured just $820 million combined. This makes it hard not to speculate that Anthropic renegotiated compute pricing with its cloud providers in 3Q26, weighing on AWS’s and GCP’s growth.

If $mrvl acquires $lite, it would likely be accepting heavy dilution to secure scarce lasers and an optical switching business that sits alongside its electrical switch silicon. Marvell’s bargaining power vis-à-vis module makers would increase, raising the question of whether the deal would be approved by regulators. $cohr will benefit from tighter laser supply and other DSP manufacturers will also benefit as Lumentum’s Cloud Light transceivers compete with the module makers that buy Marvell DSPs; some may move DSP volume toward Broadcom, Credo and others.

These are all “gray rhino” risks—obvious threats that the market continues to underappreciate. Everyone knows ChatGPT will have an impact on Google, but amid the focus on what has happened over the past year, that threat has been overlooked. Meta, too, has spent years laying the groundwork to improve its performance in historically weaker verticals.

We rarely mention another company in the title of a company’s preview. This quarter’s $GOOG preview marks only the second time we’ve done so. The first was last quarter’s APP preview, where we put Unity directly in the title. We’ll soon be doing so for a third time.

$BUG is now outperforming $SOXX at more than 75% YTD. Our cybersecurity analysts comment on the potential AI model revenue from cybersecurity and how the security market could be divided between AI labs and incumbent security platforms like $CRWD.

Bloomberg reports $MSFT will offer voluntary buyouts to ~60,000 employees. That is not 60,000 layoffs, and the difference matters for the math. A few observations: 1. Scale: Microsoft ended FY26 (June) with ~223K employees, its first YoY headcount decline in a decade (228K in FY25). 60K eligible is ~27% of the company. Eligibility is VP-level and below on select teams in the US, Canada, India and the UK. Data center, sales and marketing staff are excluded. 2. Who is likely in scope (my read): Strip out sales/marketing (~43K) and data center ops, and what remains is mostly engineering/product (R&D ~77K) and corporate functions (G&A ~14K). Microsoft is trimming the people line in exactly the areas where AI tooling is supposed to raise output per head. 3. The precedent: April's first-ever Voluntary Retirement Program covered ~8,750 US staff (age + tenure of 70+). The cost was guided at ~$900M (~$350M COGS / ~$550M opex). On the July call management said it came in lower than expected, which suggests uptake was modest. 4. Implied math (my est., 20% uptake): ~12K exits x ~$200K fully loaded cost = ~$2.4B/yr in savings, ~1.8% of FY28 opex. Severance at ~$125K/head = ~$1.5B one-time, likely booked in the Dec quarter when offers are accepted. 5. EPS impact (my est., 19% tax, ~7.4B shares): One-time charge ~-$0.16. Savings +$0.13 in FY27 (half year) and +$0.26 in FY28 (~+1%). FY27 GAAP is roughly a wash; on an ex-charge basis it is a clean +$0.13. 6. For context: Revenue per employee is already ~$1.5M (FY26 $332B / 223K). $2.4B of savings is about 4 days of FY27 capex (~$230B, my est.). This is margin hygiene, not a capex offset. 7. Caveat: Package terms aren't public, and some of the savings will likely be redeployed into AI talent and compute. Management has already guided FY27 operating margin down slightly, so the net EPS uplift could be closer to half. The real signal isn't the dollars. Microsoft is the first megacap to explicitly trade people opex for compute capex at this scale. Watch whether peers follow.

58/ Cash and funding Money raised: ~$140B in cash from 2023 through H1'26 (preferred stock + convertible notes + SAFEs, per the S-1 cash flow statements). $110B of that came in H1'26 alone, including the $65.2B Series H. Cash left at 6/30/26: $112.4B cash & equivalents + $19.3B short-term investments = ~$132B (plus $2.5B restricted cash). Additional capacity: $15B undrawn revolver (expanded Sep'26), and up to $77B of partner note facilities (Google $20B, Amazon $15B tied to compute milestones, Broadcom $42B only to pay lease rent). IPO proceeds come on top. How that compares (cash & liquid investments at 6/30/26, gross, before debt): Alphabet: $242B Anthropic: $132B Amazon: $123B Meta: $90B Microsoft: $77B Tencent: ~$75B (RMB511B) Alibaba: ~$70B (RMB475B) Anthropic is holding more cash than Amazon, Meta or Microsoft. But SpaceXAI alone runs ~$37B a year from Dec'26, against ~$780-810B of total compute commitments.

1/Anthropic’s S-1 is out. 2/ Revenue, quarter by quarter ($M) Q1'25: 376 Q2'25: 787 Q3'25: 1,382 Q4'25: 2,040 Q1'26: 4,731 Q2'26: 11,555 Q2'26 revenue was up 1,368% YoY. Compute & infra costs were up 300% over the same period. Q2'26 was the first quarter with positive GAAP operating income: $974M. 3/ How “Run-Rate Revenue” is defined Trailing-28-day consumption fees × 13 + monthly subscription fees × 12 Three caveats from the S-1: -It includes customers later found to be violating the usage policy -It is seasonal: usage dips in summer and winter holidays 4/ Revenue is recognized gross When Claude is sold through the AWS, GCP or Azure marketplaces, Anthropic is the principal. It books the full amount as revenue, and the cloud provider’s fee goes into Sales, marketing & partnerships. The marketplace share of revenue has climbed fast: 11% (2023) → 32% (2024) → 47% (2025) → 48% (H1'26) Adjust for this before comparing multiples with companies that report net. 5/ Distillation The S-1 puts a number on it: “we estimate that revenue from such customers represented approximately 5% to 7% of our revenue for the nine months ended September 30, 2026” These are customers violating the usage policy or commercial terms. Applied to ~$33B of 9M revenue, that’s roughly $1.7–2.3B. Enforcement was stepped up in Aug–Sep, and the company expects the share to decline. The S-1 says the estimate “requires judgment and may change.” 6/ Visibility is low RPO at 6/30/26: $10.8B ~75% of that is recognized in the next 12 months, so cRPO ≈ $8.1B Deferred revenue: $1.65B Q2 annualized revenue is ~$46B, so cRPO covers less than 20% of it. 80% of revenue is consumption-based, and next year’s revenue depends almost entirely on usage. 7/ The compute bill Disclosed commitments: Broadcom TPU leases: ~$161B SpaceXAI (NVIDIA GPUs): up to $121B, ~$3.1B a month from Dec’26 Google: ≥$115B, plus $125–150B more of TPU capacity in 2028–31 Amazon: ≥$110B Microsoft: ≥$45B Nscale: ~$45B Lambda: ~$41B AMD: >$20B That adds up to roughly $780–810B. 8/ The mismatch to watch Revenue is variable and consumption-driven. Compute costs are fixed, multi-year and take-or-pay. On top of that, the Broadcom leases, the credit agreement and the convertible notes include cross-default provisions. This is the single most important risk in the filing. 9/ Demand is real NRR ~900% (ex-Pro/Max) Customers paying more than $100K / $1M / $10M a year: 11,900+ / 1,900+ / ~250 No single customer was >10% of 9M’26 revenue Top 5 customers ≈ 25% of 9M revenue Q3: 79% of revenue came from outside the top 5, up from 64% in Q1 Claude Enterprise users who also use Claude Code spend ~20x more than those who don’t 10/ Dilution $77B of convertible capacity: Google $20B, Amazon $15B, Broadcom $42B, settled in shares at a trailing VWAP Google can buy up to $10B more stock at the IPO price A Public Benefit Fund already holds 5% of fully diluted shares, and can get up to +2% more for every 50% rise in the share price

As we approach AGI, more compute has to go to training, which makes ROI harder to prove, which makes the investing experience less pleasant, as our collective experience this week showed. Read our reaction to the Anthropic ARR number here: Thanks for the shoutout @SaaSletter !

$SMTC sounded confident about the next two years but conservatively guided growth in line with the industry in 2029 and 2030. More importantly, we came away impressed with the team's technical chops.

Column|Anthropic: What the New 3Q Numbers Mean Anthropic disclosed 3Q26 revenue of $16.8Bn during its just-completed pre-IPO investor day. The company had previously disclosed 2Q26 revenue of $11.5Bn and 1Q26 revenue of $4.7Bn. There is speculation that the company guided to $100Bn of ARR by year-end, but we don’t believe it has given explicit guidance. Based on the disclosures and our own tracking over time, we built a full-year ARR and revenue model for Anthropic. Detailed Report

Review| $TSM 26Q3: Revenue Guidance Beats Expectations as Expansion Weighs on Margins TSMC expressed confidence in advanced-node and long-term AI demand during its earnings call. Management said that, alongside discussions with direct customers and their downstream customers, its teams had checked US data center construction and power availability, with encouraging feedback so far. Agentic AI and open-source models are also increasing demand for computing chips. TSMC is responding by adding N3 fabs, departing from its usual practice of stopping capacity additions once a node reaches its target. Detailed Report

Review| $ASML 26Q3: Upgrades Lift Gross Margin; Capacity Expansion Awaits Further Clarity ASML’s Q3 gross margin and Q4 guidance both beat expectations. The company narrowed its 2026 revenue guidance to EUR 44–45b and raised its full-year gross margin guidance to around 56%, driven by continued strength in the upgrade business within Installed Base Management. The company stated that its plan to increase capacity by 30% in 2027, together with the further 30% increase in 2028 that had been under investigation, will still not fully meet customer demand, and it is now looking into expanding 2028 capacity beyond 30%. Customers have expressed interest in placing orders for 2029 and 2030, although ASML has not yet accepted those orders. We raise our 2026/27/28 EPS forecasts to EUR 42.5/61.4/76.9. Detailed Report

$ORCL and $CRWV face a repricing of credit risk, along with SoftBank in Japan. Higher credit risk premiums → potentially higher costs on new financing → pressure on neoclouds’ purchasing power → a possible downgrade in AI hardware order growth. $ORCL: 5Y CDS widened 26bp overnight to a record 288bp. $CRWV: CDS broke through 1,000bp, and its 2032 bonds fell to 77 cents on the dollar. CDS on Amazon, Alphabet, Meta and NVIDIA also hit record highs, though their spreads remain well below Oracle’s. Our AI Capex Credit Watch:

About 120-140 people attended our intimate @FundaAI happy hour. I was blown away by some of the feedback we received from clients and attendees. A special thanks to Funda for making this possible and to all attendees, and apologies if I wasn't able to chat with all of y'all.

A robot rang the closing bell at Nasdaq yesterday. The day before, the CEO of Unitree said his conviction in AI development grows every day. Based on strong GPT-6 test results in varied tasks, he believes GPT-7 could fully resolve embodied AI challenges. We think robotics may become the next focus of competition, driving a second wave of compute demand. Read more here:

Research|META: Muse Wonder Returns Meta to the Frontier as Compute Tightens Personal agents expand Meta’s opportunity as training and inference demand put more pressure on compute capacity On October 12, Meta Superintelligence Labs (MSL) released Muse Wonder and Muse Wonder Full, the models previously referred to in reports as Watermelon. These are Meta’s first frontier models since Llama 4, arriving 14 months after Alexandr Wang reorganized the company’s AI teams in August 2025. Both were trained on Meta’s Prometheus cluster in Ohio, support text, image, audio, and video inputs, and have a 1M-token context window. Neither model’s weights are publicly available. Wonder is available through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens. Wonder Full is available only to Muse Code subscribers on the $50 and $200 monthly plans, with no API access. Our channel checks indicate that Wonder has roughly 3T total parameters, making it smaller than both Astra and Fable. We believe Meta is back among the leading frontier AI labs, and we expect it to stay there. In our view, a year ago that group consisted mainly of OpenAI and Anthropic. Google rejoined this year, followed by Meta, and a four-way contest is taking shape. The hiring spree in the summer of 2025 is paying off: backed by substantial investment in compute, the new team brought Meta back to the frontier in just over a year. Compute remains the biggest bottleneck. According to media reports, in July Meta was preparing to rent out surplus GPUs and was in talks with Anthropic; by October, its own needs had forced it to shelve the plan. Anthropic, meanwhile, secured new capacity from other suppliers and raised usage limits when it launched Opus 5.5. We expect the shortage to ease only once Vera Rubin and TPU 8t ship in volume in 2027. We also expect Meta’s 2027 capex to rise again. On open-weight models, we maintain our September view that the gap between the leading open-weight models and the top closed models will widen again to 6 to 12 months. Detailed Report

The $zs Investor Day last week did not resolve the near-term growth debate (no guidance change), but materially strengthened the platformization thesis. The newer businesses are becoming increasingly measurable: Branch and Cloud reached roughly $250M of ARR, Data Security about $410M, while Security for AI is emerging from a smaller base with strong bookings and pipeline growth. At the same time, Z-Flex is helping convert Zscaler’s broader product portfolio into higher wallet share, deeper multi-product adoption and stronger retention. The strategic direction is therefore becoming clearer: Zscaler is evolving beyond its core ZIA/ZPA franchise into a broader security platform spanning users, workloads, branches, data, AI and SecOps. The key remaining question is whether these newer engines can reaccelerate consolidated ARR growth from ~17% (FY27) back toward 20%+.