SemiAnalysis Datacenter Energy & Anthropic Disruption Investment Thesis

Source: Ep. 004 — The Impact of AI Datacenters On Consumer Power Costs, SemiAnalysis Weekly (Jordan Nanos, Doug O'Laughlin, Jeremie Eliahou Ontiveros), March 5, 2026.

The Framework: Grid Design, Not AI Headlines, Sets Consumer Power Prices

Jeremie Eliahou Ontiveros (Head of Datacenter & Energy Infrastructure Research at SemiAnalysis) argues the popular narrative — AI data centers are driving consumer electricity bills higher — is directionally true but massively over-indexed on the wrong mechanism. The organizing split is PJM (auction, backward-looking) vs ERCOT (real-time, market-driven) plus whether hyperscalers pay their own grid upgrade costs via customized tariffs.

Factor	PJM (VA, OH, NJ)	ERCOT (Texas)	Investment Read
Price setting	Annual auction, 2 years ahead	Real-time market signals	ERCOT adapts faster; PJM lags
Demand forecast	Undercounts data centers	Market participants price ahead	SemiAnalysis model vs PJM error
Supply accounting	Regulatory methodology cuts ~14 GW	Market-driven additions	Coal retirements + accreditation reform
Datacenter connection	Mostly grid-tied today	Mix; behind-the-meter growing	Custom tariffs + Oracle-style deals
Consumer bill risk	Transmission ROE pass-through if overbuilt	Lower if tariffs isolate load	Hyperscaler prepayment = mitigation

Investment Thesis #1: PJM Consumer Bill Spikes Are a Market-Design Failure, Not Pure AI Fault

Jeremie’s 7,000-word SemiAnalysis article responds to Trump-era legislation rhetoric blaming AI for rising electricity bills. PJM-region consumers face 15–20% bill increases vs. two years ago — but the driver is how PJM runs its capacity auction, not datacenter load alone.

"It's actually way more driven by market design... This market was not designed for an environment where load is growing."

PJM forecasts demand poorly (missing data center load SemiAnalysis tracks in its model), runs a once-a-year auction priced two years forward, and regulatory supply accounting — not capitalism — sets marginal generator prices. ERCOT lets participants react in real time and place generation ahead of demand. After 20 years of zero US load growth, PJM’s communist-vs-capitalism design (Jeremie’s framing) cannot handle AI-era load growth without violent price swings.

Trigger: PJM capacity auction clearing prices; PJM demand forecast revisions vs SemiAnalysis datacenter model; state-level legislation blaming AI without tariff reform.

Names: No direct equity call — thesis is regulatory/market structure. Indirect: datacenter-heavy utilities in PJM states benefit from customized tariff negotiations.

Investment Thesis #2: Hyperscaler Grid Commitments (Oracle Michigan) Are the Template — Pass-Through Risk Falls on Oracles, Not Consumers

The worst consumer outcome: overbuilt transmission assets with low utilization — regulated utilities earn guaranteed ROE and pass costs to retail ratepayers. The fix: customized tariffs where hyperscalers prepay grid upgrades, batteries, and long-term demand charges.

"Oracle signed... a $4 billion commitment with the utility... $2 billion of batteries... a 20-year agreement... with a minimum demand charge."

Oracle’s Michigan Stargate site (1 GW IT / 1.4 GW gross peak, OpenAI tenant via Correlated Digital) includes a $4B utility commitment — ~$2B for grid batteries, plus 15–20 year power pricing with minimum demand charges even if load under-materializes. Negotiations began ~2 years ago; Oracle signed ~October 2025; operational target 2027, full campus by 2028. Total timeline ~5 years — debunking the idea that grid deals happen in six months.

Most datacenters coming online today remain grid-connected under legacy rules; behind-the-meter buildout accelerates late 2026–2027.

Trigger: Oracle utility filings in Michigan; similar tariff announcements from other hyperscalers; battery storage capex tied to datacenter sites.

Names: Oracle (ORCL) — explicit $4B grid commitment model for Stargate; datacenter developers and utilities signing parallel deals.

Investment Thesis #3: Coal Retirements Pause as Prices Rise — Dispatchable Generation Gets a Second Life

PJM supply fell ~40 GW (~20%) in four years — partly real coal retirements, partly thermal accreditation reform after 2021–2022 winter storms showed plants fail in clusters (41% / 14 GW of the decline is methodology, not physical removal). Federal policy is delaying coal plant retirements (Q4 2026 cohort). Old plants are uneconomic at low prices but become viable when capacity prices spike.

"You have to balance consumer expectations of prices staying low, but... to incentivize new generation to come online, you need prices to actually be high."

Behind-the-meter generation sidesteps grid queue delays — hyperscalers build what they need without waiting for PJM reform.

Trigger: DOE/FERC coal retirement delay orders; PJM capacity auction prices sustaining elevated clears; behind-the-meter gas/coal announcements at datacenter campuses.

Names: GE Vernova (GEV) and gas-turbine oligopoly names implied by dispatchable generation demand (not named explicitly in episode); behind-the-meter power developers.

Investment Thesis #4: Anthropic ARR Crossover — Model Labs Eat the Application Layer

Doug O'Laughlin argues Anthropic crossed OpenAI’s total ARR in March 2026 (February exit run-rate) — roughly a 2–3 month lead — with ARR doubling in two months (~~$10B added). Reported Claude Code ARR (~~$2.5B) is materially understated because SemiAnalysis spends 90–95% on API, which Anthropic does not attribute to Claude Code.

"The idea that Anthropic's going to come for Salesforce... seems right to me. The model companies will become the everything companies."

Claude for finance, legal, and security launched sequentially and tanked SaaS stocks. Doug compares to Moore’s Law: when general models improve every 3–6 months, specialization loses — like CPUs crushing custom chips until Moore’s Law ended. Coding is upstream of all information work; Anthropic’s RL flywheel requires coding-quality traces, not chat votes.

Department of War headlines may have boosted app downloads (#1 above ChatGPT) but Doug attributes revenue acceleration to enterprise API adoption, not political sympathy downloads.

Trigger: Anthropic ARR disclosures; Cursor/Windsurf/Vercel API volume; SaaS seat churn at Salesforce-like vendors; Claude enterprise SKU launches.

Names: Anthropic (private) — revenue leader and vertical SaaS disrupter; Oracle (ORCL) — Stargate tenant exposure to OpenAI competitive position.

Bearish (Key Risks only): Salesforce (CRM) — enterprises removing CRM building custom stacks on Claude API.

Investment Thesis #5: Chinese AI Demand (ByteDance/Seedance) Is an Under-Appreciated GPU Demand Vector

SemiAnalysis notes ByteDance Seedance has a 3-month paid queue (14K of 30K in paid tier) — users cannot pay to skip the line. Chinese domestic compute cannot support demand; usage VPNs through Japan/Korea — at one point Korea+Japan were ~15% of Anthropic revenue, largely Chinese labs distilling Claude. Doug: if Eastern labs had Western compute parity, they would be ahead; distillation (GLM 4.7 on Claude) already competitive.

This leaks into Singapore, Europe, and global GPU demand — a demand vector Western CapEx models may undercount.

Trigger: ByteDance/Seedance queue metrics; Anthropic APAC revenue attribution; Singapore/Malaysia datacenter lease announcements from Chinese tenants.

Names: Alphabet (GOOG), Nvidia (NVDA) — indirect beneficiaries of ex-China compute demand; no direct ByteDance equity (private).

Investment Thesis #6: HBM Long Thesis Intact — SRAM/LPU Scare Is Investor Noise

Doug dismisses the narrative that SRAM/LPUs will displace HBM as investors rotate scared out of memory longs. GTC LPU session hype (Groq) is real investor attention, but Doug calls it "the dumbest narrative I've heard recently" — people are too long HBM and looking for an exit story.

Trigger: GTC LPU benchmarks vs HBM systems; Groq deployment metrics failing to displace HBM attach rates.

Names: Micron (MU), SK Hynix — HBM thesis unchanged per SemiAnalysis panel.

The Ecosystem Map

Energy / grid:

SemiAnalysis datacenter-energy model (Jeremie) — PJM forecast error, tariff tracking
Oracle Michigan: $4B utility deal, $2B batteries, 20-year demand charge
Behind-the-meter acceleration: late 2026–2027
Coal retirement delays at federal level

AI software disruption:

Anthropic: ARR crossover, API > Claude Code reported revenue
Tailscale: CoreWeave cluster auth; accelerating in agentic era (Doug)
Rippling: accelerates with AI workforce tooling (mentioned)
Obsidian: Doug "graduating" to terminal/vector DB workflows

Chinese demand:

ByteDance Seedance queue; distillation from Claude; VPN demand through APAC

Key Risks

PJM reform succeeds: Better forecasting and real-time pricing could collapse capacity prices — reducing generator and coal-continuation economics.
Anthropic DoW supply-chain designation: If vendors are forced to cut Anthropic compute, revenue crossover thesis stalls.
OpenAI re-acceleration: Doug acknowledges "until OpenAI comes back" — Codex 5.3 and distribution could narrow ARR gap.
Oracle Stargate demand fails: Minimum demand charges protect utilities, but Oracle bears committed cost if load slips.
Transmission overbuild without tariffs: If hyperscalers connect without customized deals, consumer pass-through risk rises — political backlash could slow datacenter permits.
SaaS incumbents respond: If Salesforce/Microsoft ship competitive agent layers on equal model cadence, application-layer disruption slows.
Chinese export/compute restrictions: Distillation crackdowns could shift but not eliminate APAC API demand patterns.

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	Oracle (ORCL)	$4B Michigan grid commitment template; Stargate 1 GW; 20-yr demand charge	Utility filing; Oct 2025 deal; 2027 ops target
1	Anthropic (private)	ARR crossed OpenAI Mar 2026; doubled in 2 months; API undercounted	Enterprise Claude for finance/legal/security launches
2	Micron (MU) / SK Hynix	HBM long intact; SRAM/LPU fear is noise per Doug	GTC narrative vs actual HBM attach
2	GE Vernova (GEV)	Dispatchable gen economic at high PJM clears; coal delays	PJM capacity prices; coal retirement delays
3	Tailscale (private)	Agent-era infra; CoreWeave cluster auth pattern	Enterprise VPN/zero-trust adoption with neoclouds
3	Nvidia (NVDA)	Chinese demand leaks globally; distillation drives API spend	APAC API revenue; GPU demand from ByteDance queue

Monitoring Checklist

Bottom Line

Consumer bill anger is PJM's fault more than AI's: 15–20% hikes trace to auction design and supply methodology, not datacenter load alone — customized tariffs are the investable fix.
Oracle's $4B Michigan deal is the grid financing blueprint: Batteries + 20-year demand charges shift transmission risk off consumers onto hyperscalers willing to prepay.
Anthropic just passed OpenAI on ARR with API spend the market undercounts: Model labs are eating the application layer — Salesforce-style SaaS is on the wrong side of the cadence.
Chinese AI demand (Seedance queues, Claude distillation) leaks into global GPU/API spend: A demand vector Western CapEx models under-appreciate.
Ignore the SRAM-kills-HBM panic: SemiAnalysis panel treats LPU hype as overcrowded-memory exit narrative, not a fundamental HBM thesis break.