Akamai + Anthropic: The $11.6B CPU Bet That Rewrites AKAM

A dot-com CDN just signed the largest contract in its history — 5x its cloud run-rate — to run agentic AI on CPUs. The re-rating is real, but the spike is not the entry.

Prices: AKAM $117.27 (+6.2% intraday Sep 25, high $129.57, prev close $110.41) · 10Y 5.17–5.19% (19-yr high) · S&P 500 7,704 · BTC $83,623 · Data via financial handler + Akamai 8-K/press release Sep 24, 2026. 中文版 → Chinese edition

1. What happened is not a cloud contract — it is a balance-sheet pivot

Akamai signed $11.6B over 7 years to supply Anthropic with dedicated CPU cloud + managed services, with a contractual path to ~$20B. The 8-K (filed Sep 24, event Sep 18) makes this Projects 2 & 3 under a Master Services Agreement dated May 5, 2026 — i.e., this relationship scaled 20x in four months.

Read the Akamai press release (Sep 24, 2026) and the 8-K Item 1.01 together, because the press release omits what the filing discloses:

• Scale: $11.6B committed, 7-year terms from service-start dates. That is ~$1.66B/year run-rate — 4.2x Akamai's entire 2025 CIS revenue ($314M) and 16.8x Q2 2026 CIS annualized ($99M × 4).
• Expansion option: +$9B on mutually agreed terms = ~$20B total. Each extra $3B vests ~1% of Akamai to Anthropic.
• Warrant: 387,051 Series B preferred → 7,741,020 common shares (~5% outstanding) at $111.33 (30-day VWAP × 20). ~2% vests on first Project 3 payment; rest on expansion. Cash-settled exercise, 7-year life, non-transferable except to Anthropic subs.
• Cost to serve: ~$5.5B total capex, including ~$1.7B in 2026 alone to pre-buy memory. Executed via Jabil $1.7B memory build request (consignment model) and a 7-year Lenovo hardware SOW (Sep 23).
• No 2026 revenue lift: Akamai explicitly guides no impact to 2026 revenue. Management whispered $150–$300M in 2027 (per Morgan Stanley note).

Consequence: this is not incremental bookings. With $14.6B in CIS commitments signed YTD (vs. $2.8B disclosed at Q2), Akamai just pre-sold ~3 years of its current total company revenue ($4.45–$4.53B guide) in one vertical.

2. Akamai was a melting ice cube with a fast-growing sidecar — until yesterday

Before the deal, Akamai was 55% security, 36% declining delivery, 9% hypergrowth cloud. Q2 2026 (reported Aug 6) tells the story: total $1.10B +5% YoY, Security $604M +10%, Delivery $396M –6%, CIS $99M +39% YoY, GAAP op margin collapsed to 7% (–8pts) and non-GAAP to 25% (–5pts) on cloud build.

Full-year guide was pedestrian: $4.445–$4.530B (+6–8%), non-GAAP EPS $6.40–$7.05, op margin 25–26%. At $110 pre-deal, AKAM traded ~15.5x 2026 non-GAAP EPS midpoint — a value multiple for 5.9% trailing revenue growth, 40.2x trailing GAAP P/E, ROE 8.5%. The 52-week range ($70.82–$165.45) shows how violently the market has swung between “CDN in decline” and “edge-AI option.”

My opinion: the bear case was never wrong on delivery — it was early on obsolescence. Delivery still shrinks mid-single digits, but security (10% growth, 57.5% gross margin company-wide) pays the bills while CIS scales. The Anthropic deal does not fix delivery; it makes delivery irrelevant — if executed. Piper Sandler’s line is exactly right: CIS overtakes Security by late 2028 at this pace, flipping AKAM from value to hypergrowth.

3. Why CPUs — and why Akamai — is the real signal

Anthropic chose distributed CPUs, not GPUs, for agentic inference — and that validates a second AI hardware stack. Per CNBC’s roundup of Wall Street notes, Oppenheimer sizes the deal at ~77MW at ~$22M annual revenue/MW (~$1.7B run-rate). CPUs handle coordinated, low-latency, memory-heavy agent tasks (voice agents, robotics, real-time video, world simulation — all named by DA Davidson from Akamai’s call) better and cheaper than GPU clusters.

BofA’s instant reaction — hiking AMD’s target to $720 because the deal “further supports CPU importance in the agentic era” — is the tell. My read, as opinion: NVIDIA trains, AMD + Intel + Akamai’s edge serve. Akamai’s 4,000+ PoPs (core-to-edge continuum) let Anthropic run inference near users/agents without building new mega-campuses — critical when Oracle just declared force majeure on a data-center build and Blue Owl’s Project Jupiter slipped a year on power constraints.

Downstream consequences that matter for your portfolio:

• Memory is the choke: $1.7B memory pre-buy via Jabil explains why MU ($1,078, –0.2%) and SNDK stay bid into Sep 30 earnings — DRAM/NAND pricing power just got a 7-year bid. A Hormuz reopening that drops Brent from ~$106 would help freight/power costs but not memory tightness.
• Neocloud repricing: CoreWeave ($87.80, –2.6%) and Nebius ($237, –2.6%) fell today while AKAM spiked. If 77MW prices at $22M/MW with 30% EBIT (DA Davidson), Akamai just set a public comp that undercuts GPU-neocloud $/MW and forces their next contracts to prove GPU premium.

4. The math works — if you believe 30% margins on $1.66B

At full run-rate, the Anthropic base deal alone is worth roughly $4–$5 of EPS — doubling Akamai’s earnings power. Here is the checkable arithmetic (see Appendix B for assumptions):

• Revenue: $11.6B / 7 = $1.657B/year. At $20B total, $2.86B/year.
• EBIT: DA Davidson’s 30% on CIS commitments → ~$497M annual EBIT base, ~$857M if expanded.
• After 19% non-GAAP tax: ~$403M NOPAT base. On ~150M diluted shares = ~$2.68/share incremental.
• Base business: $6.70 midpoint 2026 non-GAAP EPS, growing ~12–16% (MS) → ~$7.50–$8.00 standalone by 2028 before Anthropic ramp.
• Combined 2028 power: ~$10–$11 EPS. At a 15–17x infrastructure multiple = $150–$187 — bracketing today’s $158–$185 Street targets.

Capex intensity is the load-bearing assumption: $5.5B / $11.6B = 47% capex-to-contract, or ~$70M/MW (Oppenheimer). That is financeable — Akamai held $4.62B cash/securities at Jun 30 and generates ~30% OCF margin ($326M in Q2) — but it front-loads spend into a 5.2% yield curve. Every 100bps on $3B of incremental debt is ~$30M pre-tax, or ~$0.16/share. In 2022 parallel conditions (BofA’s warning today), that math killed unprofitable growth. Akamai is profitable, but free cash flow goes negative in 2026–27 before inflecting in 2028.

My judgment: fair value $165 base (15x 2028 $11), $185 bull (expanded $20B at 70% probability-adjusted), $95 bear (execution slip + customer concentration discount). Do not pay $129 for it today — the stock tagged $129.57 intraday (+17%) then faded to $117 as rates re-spiked. The warrant strike $111.33 is your gravitational center.

5. Risks Wall Street skimmed over in the first 12 hours

One customer, one architecture, one memory cycle — funded at 19-year-high rates.

• Concentration: Anthropic = ~37% of forward revenue at run-rate. Termination triggers are narrow (material uncured breach, outage, change-of-control to direct competitor), but Anthropic can walk project-by-project on outage. A Pentagon “supply-chain risk” ruling upheld today (appeals court 2-1 vs. Anthropic) adds headline overhang to Anthropic’s government-adjacent workloads.
• Execution + supply chain: $1.7B memory via Jabil consignment — Akamai pays on receipt, Jabil repurchases at cost on use. If Lenovo/Jabil slip or DRAM prices spike, Akamai eats carrying costs. Q2 GAAP op margin already –8pts on early build; more pain before gain.
• Dilution: 5% warrant at $111.33 caps upside mechanically — every $10 above strike on 7.7M shares is ~$77M of value transfer. Manageable, but real.
• Rate regime: 10Y at 5.19% (Trading Economics Sep 25), MOVE 104 (highest since March) vs. VIX 15.2 complacent. Schwab notes 66% odds of an October hike under Chair Warsh. Goldman’s cold check today — hyperscalers need ~$300B AI revenue in “next few years” to break even — applies to Anthropic too. If Anthropic’s $14B revenue run-rate stalls, Akamai’s $5.5B capex is stranded.
• Technical: AKAM is still –29% below its May $165.45 high. $128–$130 (today’s spike high) is now resistance; $110 (prev close / warrant zone) is intraday support.

6. Trade: buy the edge-cloud repricing, not the headline spike

Chasing +17% into a bond selloff is how you turn a great contract into a bad trade. Wait for the rate-driven fade; Akamai always fades day-1 AI spikes (it did after the $600M robotics deal in August).

Bull–base–bear (probability-weighted, 12-month):

• Bull 25% ($210): +$9B expansion triggered by mid-2027, second frontier lab signs, CIS >50% growth, 10Y back <4.8%. AKAM rerates to 18x $11.50 = $207.
• Base 50% ($165): $11.6B delivers on schedule, $150–$300M in 2027, margins trough then recover, 10Y 5.0–5.3%. 15x $11 = $165.
• Bear 25% ($95): Memory overruns, 10Y >5.5% forces debt, Anthropic pauses expansion, delivery drag persists. 12x $8 = $96.
• Expected value ≈ $156, ~33% above $117 — positive skew if you enter in the zone, negative if you chase $129.

7. What would prove me wrong — and what to watch next

I am wrong if Akamai cannot convert watts to dollars. Specifically: (1) Q4 2026 or Q1 2027 CIS revenue fails to accelerate past $130M/quarter (implies <$150M 2027 Anthropic contribution tracking); (2) capex ex-memory exceeds $3.8B residual without MW disclosure; (3) Anthropic does not make first Project 3 payment (warrant 40% fails to vest) by Q1 2027; (4) daily close below $98. Any one triggers exit, no second chances.

Catalyst calendar: Micron earnings Sep 30 (memory pricing read-through); PCE + JOLTS next week (rate path for funding); Akamai Q3 print early November (watch CIS guide ex-Anthropic); Lenovo/Jabil delivery milestones in Q4 10-Q exhibit filing (the full MSA text lands with Sep 30 10-Q).

Positioning note vs. open book: this complements EQ-ORCL-1 (fade Oracle on force-majeure/build risk) — Akamai is the distributed, asset-light mirror to Oracle’s centralized mega-build. It does not conflict with MACRO-5PCT-1 (fade duration above 5.10%); if 10Y breaks sustainably below 5.05%, add to AKAM — cheaper funding steepens skew to bull.