Anthropic's $11.6B Akamai deal, and a 5% option

• FACT (CONFIRMED — 8-K, EDGAR, retrieved directly): On September 18, 2026, Anthropic, PBC and Akamai signed Project Plan 2 and Project Plan 3 under their MSA dated May 5, 2026 for dedicated cloud computing capacity and related managed support services. Anthropic committed to pay approximately $11.6 billion in aggregate, each plan with an initial seven-year term from its service start date, subject to delivery and service-availability requirements. Akamai determined the MSA "is no longer immaterial in amount or significance." Termination: Akamai may terminate for Anthropic's uncured breach; Anthropic may terminate upon a material uncured Akamai breach or upon a change of control of Akamai in favor of a direct Anthropic competitor; either party on bankruptcy; Anthropic may also terminate each Project Plan upon notice of a material outage (conditions apply). The MSA full text will be filed as a 10-Q exhibit.
• FACT (CONFIRMED — 8-K + press release): On the same day (Sept 18, the "Issue Date"), Akamai issued Anthropic a warrant to purchase up to 387,051 shares of Series B Non-Voting Convertible Preferred Stock at $2,226.60 per preferred share — the 30-trading-day VWAP of Akamai common before the Issue Date, multiplied by the initial 20:1 conversion rate → 7,741,020 common shares as-converted ≈ 5% of shares outstanding, common-equivalent price $111.33. Vesting: four tranches — 40% of warrant shares (≈3.096M common, ~2% of shares outstanding) upon Anthropic's first payment under Project Plan 3; three successive 20% tranches (≈1% each) upon each additional $3.0B of contractual value committed. Cash settlement required (total ~$861.8M to exercise all vested shares — computed and cross-verified in labs/S05.md). Exercisable until Sept 18, 2033. Non-voting (except DGCL rights), dividend-equivalent, $0.01 liquidation preference, converts to common only on transfer outside Anthropic's wholly-owned family; Section 4(a)(2) unregistered issuance; transfer-restricted.
• FACT (CONFIRMED — press release + transcript): The deal is framed as supporting Anthropic's "accelerating CPU workload demands" on "Akamai Cloud's distributed AI infrastructure and software." Expansion: "potential expansion of the relationship by up to an additional $9 billion… total potential commitment of approximately $20 billion." It adds to >$2.8B of multi-year CIS commitments previously announced this year → ~$14.4B YTD signed TCV (transcript). Total related capex ~$5.5B; +$1.7B 2026 capex to pre-purchase components including memory; no change to 2026 revenue guidance.
• COMPANY CLAIM (management guidance — investor call, Sept 24, Ed McGowan CFO): Revenue: none in 2026; begins 2H 2027; $150–300M FY2027; fully ramped by year-end 2028 at ~$1.7B annualized run-rate for the rest of the take-or-pay term ("guaranteed payment upon delivery"). The warrant's fair value is deducted from revenue ratably over the contract and is inside the $11.6B figure; ASC 606 straight-lines escalators. Capex timing: ~$1.7B Q4-2026 (memory pre-buy), ~$3.1B 2027, ~$0.7B 2028; 7-year depreciation assumption; no contractual hardware-refresh obligation; portfolio 95–105 MW across ~$14.4B TCV (not deal-specific power); colo ~$3–4M/MW; temporary margin pressure during 60–90-day site ramps; funding from $4.6B cash + $1B revolver; treasury-stock-method dilution. Warrant value: grant-date value "a little over $150 million" vs the $11.6B deal, and "less than $400 million" if the full 5% vests on a $20B relationship — "very, very small compared to the overall revenue."
• FACT (CONFIRMED — 8-K): Same package: Lenovo MPSA (entered Sept 23, 2026; hardware, software and services; 3-year initial term, 7-year SOW No. 1, undisclosed value) and Jabil Build Request (Sept 24, 2026: authorization to buy ~$1.7B of memory components under the 2019 Jabil MSA; Jabil holds them in consignment as bailee and repurchases at cost as consumed).
• INDEPENDENT EVIDENCE (market reaction, multi-source): AKAM closed Thursday Sept 24 at $110.41, then: +22% extended trading (Reuters), "as much as 17% to $129.60" (Bloomberg via TNW), 16% (CNA syndication of Reuters); Friday premarket +21.3% (Yahoo/finviz); Friday open $125.23, peak +16.4% to $128.46, faded to +8% midday ($119.59) (CNBC/Awesome Agents), official close $113.94, +3.20% (stockanalysis.com close data; Markets Insider quote widget concurs; TradersUnion printed $114.69/+3.85% — minor feed discrepancy, recorded).
• CONTEXT (INDEPENDENTLY VERIFIED reporting): TechCrunch: this is >6× the $1.8B seven-year deal Bloomberg reported in May 2026 (same MSA lineage), and the first time Akamai has attached a warrant to a cloud deal (Bloomberg). Reuters: Anthropic is "IPO-bound" and last month agreed to spend $45B renting AI cloud from Nscale's West Virginia campus (Reuters, Aug 2026). TNW/Bloomberg: Anthropic's 2026 compute run includes a $10B Volta deal (Aug), reported role as the customer behind Rum Group's $13.7B contract (The Information, Sept), first Australian data-center lease, chips from Google and SpaceX (Bloomberg) — these are press-reported, not Anthropic-confirmed. eMarketer's Jacob Bourne (Reuters): infrastructure investment "will continue to expand to support increasingly compute-intensive agentic workloads"; the stake "can be seen as a vote of confidence in the durability of AI-driven cloud demand."
• COMPANY BEHAVIOR NOTE: Anthropic was not quoted in Akamai's announcement (TNW; verified: no Anthropic newsroom post found). CEO Dario Amodei's relevant public position (via TechCrunch, citing NYT, Dec 2025): Anthropic does not participate in circular investor-supplier deals "at the same scale as some other players."

• A frontier lab bought CPU-scale infrastructure at hyperscaler-fleet dollar levels (FACT of the purchase; INTERPRETATION of significance). Until this week the AI-infra arms race was GPU-headline-driven (training campuses, GPU clouds). An $11.6B seven-year CPU-first contract — 6× Akamai's May deal, ~$2.2B ARR from Akamai's YTD CIS portfolio — makes general-purpose compute for agent orchestration, serving and adjacent workloads a named, priced, contracted layer of AI strategy. Neither company disclosed what the CPUs will run (TechCrunch noted the gap), but Akamai's own pitch maps it to agent-era work: conversational voice agents, sandboxed code execution, robotics/physical AI, real-time video intelligence, agentic DevOps (Leighton, call).
• The financing structure of AI deals flipped direction (FACT of structure; INDEPENDENTLY VERIFIED novelty via Bloomberg). The "circular AI deal" pattern has been suppliers buying equity in their hot customers (NVIDIA→OpenAI, AMD→OpenAI milestone warrants, Amazon/Google→Anthropic while selling compute). Here the supplier issued the customer a warrant — Akamai giving its biggest buyer a staged ~5% equity option, vesting 1% per incremental $3B of spend. TechCrunch: "the warrant flips the more common pattern"; Bloomberg: Akamai's first warrant-attached cloud deal; Leighton (via Bloomberg): "It helps bring the companies together." Whether this becomes the template for second-tier clouds courting labs is now an open market-design question with a live example.
• Akamai's identity changed in one filing: from CDN relic to contracted AI-infrastructure balance sheet (FACT of what was disclosed; INTERPRETATION of market meaning). YTD signed CIS TCV ~$14.4B against a 2025 revenue base that grew 5% (Morgan Stanley note via CNBC); management guided the Anthropic contract alone from zero to ~$1.7B/year by 2028 — roughly a ~16% growth-rate re-rating of the whole company (MS estimate). Piper Sandler: "from a 'value' asset to a 'hypergrowth' asset." The $5.5B capex + $1.7B memory pre-buy + Lenovo/Jabil agreements show Akamai financing the buildout from operating cash ($4.6B) + revolver, i.e., staking its balance sheet on the ramp.
• The memory market got a new $1.7B buyer committing through 2026-Q4 (FACT of authorization; INTERPRETATION of supply effect). An unusual disclosure in its own right: Akamai pre-purchasing $1.7B of memory components via Jabil consignment "to secure critical supply chain" amid industry-wide memory tightness — enterprise-grade evidence that component scarcity, not just datacenter space, is now the binding constraint on AI buildouts.
• Governance terms reveal who holds leverage (FACT from 8-K; INTERPRETATION of the read). Anthropic negotiated a change-of-control termination right if Akamai is acquired by a direct Anthropic competitor, plus per-plan termination on material outages — customer-protective clauses that tell you what Anthropic fears (being hostage to an infra vendor that gets bought by OpenAI's or Google's parent) and confirms the contract is not the unconditional take-or-pay some coverage shorthand implied (TechCrunch: "isn't ironclad").

Contract mechanics (CONFIRMED from 8-K/press release/transcript):

Why CPUs, why distributed (as told by primary sources + press): Akamai's pitch is a "continuum of compute from core to edge" over 4,000+ PoPs — the layer that runs agent orchestration, code execution/sandboxes, routing, media/security services and CPU-side serving rather than tensor math. CNBC: "Networks of CPUs are generally better for the types of coordinated computing tasks required for agentic AI, many technologists say." CFO: CPU deals yield more revenue per MW than GPU (denser silicon per watt of useful compute) and the ~1:3 ARR:capex on this deal (vs the ~0.5:1 portfolio norm) reflects power efficiency + seven-year scale pricing (smaller customers pay much higher per-unit yields). What Anthropic actually runs on it: not disclosed by either party (TechCrunch explicitly flagged).

• It prices the agentic-inference era in hard dollars (INDEPENDENT EVIDENCE: SEC-filed contract; INTERPRETATION: what it signals). $11.6B for seven years of CPU/distributed capacity — ~$1.7B/year run-rate against Akamai's ~$4.3B 2025-scale revenue — is the largest public-market statement yet that AI's cost stack is not only accelerator chips. For anyone modeling AI economics: agent platforms need orchestration, execution sandboxes, retrieval, routing, and CPU-heavy serving at scale, and that demand is now contracted through 2033 by the lab growing fastest (Reuters: "one of the fastest-growing AI companies"; Dario Amodei has publicly cited 80x annualized growth earlier in 2026 — pre-window context).
• The equity-warrant mechanism re-distributes who captures AI upside (FACT of structure; INDEPENDENTLY VERIFIED novelty via Bloomberg/TechCrunch; INTERPRETATION of consequences). Suppliers taking equity in labs is the known circular-deal critique. Labs taking equity in suppliers — vesting per incremental spend — is the mirror image: Anthropic can hedge its infra dependence by owning the vendor, and Akamai pays for demand certainty in paper instead of margin. The CFO's framing ("value very, very small compared to overall revenue… helps align interests") is a COMPANY CLAIM optimizing the narrative; the strategic read (Anthropic building a compute-supplier portfolio partly as an investment book before a rumored IPO — "IPO-bound," Reuters) is INTERPRETATION. Expect procurement teams and other labs to price this structure into every large cloud negotiation from now on.
• Second-tier clouds are viable AI suppliers at frontier scale, but the cost of belief is balance-sheet risk (INTERPRETATION on CONFIRMED facts). Akamai staked $4.6B of cash + a revolver on a take-or-pay ramp with disclosed 60–90-day site margin pressure; every dollar of the story depends on execution from 2H-2027. The +3.2% official Friday close (after a 20%+ paper spike) shows sophisticated money fading the spike — the market's way of saying "we believe the backlog, we're waiting for the ramp."
• The supply-chain arms race now has component pre-purchasing as a public disclosure event (CONFIRMED). A $1.7B memory pre-buy through a contract manufacturer's consignment, inside a customer-deal 8-K, is the clearest filed evidence yet that memory procurement is a strategic chokepoint for AI infrastructure. Enterprises buying AI services should expect supplier cost volatility to persist into 2027; investors in memory/storage names just watched a non-hyperscaler compete for allocation.
• Anthropic's silence is itself information (FACT: no statement found; INTERPRETATION). Akamai told the whole story; the lab said nothing. Anthropic has consistently let partners announce compute deals (May precedent). Combined with Amodei's December NYT distancing from circular deals ("not… at the same scale as some other players"), the company appears to want optionality optics without endorsement — and the 8-K's change-of-control clause reveals the strategy: secure capacity, but never let a supplier get captured by a rival model lab.

• For Anthropic: contracted, non-hyperscaler CPU capacity for agent-era workloads through 2033, with a potential ~5% equity call on its supplier and an exit door if Akamai is ever acquired by a direct competitor — a template for de-risking compute dependence while keeping upside alignment.
• For Akamai: a re-rating anchor — $14.4B YTD contracted CIS TCV, a guided path from ~5% to mid-teens revenue growth, a validated "distributed CPU cloud" positioning, and its first warrant-attached customer structure as a competitive weapon for courting other labs.
• For the market: a visible, filed price discovery on CPU-for-agents economics (~$22M revenue/MW portfolio, ~1:3 ARR:capex, $3–4M/MW colo) that analysts can underwrite — Oppenheimer's ~77 MW estimate for this deal is now a computable number, not folklore.
• For other AI vendors: the second-tier-cloud path to frontier contracts is demonstrated with terms anyone can copy — capacity commitments + supply pre-buys + customer warrants — which likely raises the floor for what mid-tier clouds (and what labs) will accept in 2027 negotiations.
• For memory/component suppliers: a public signal that compute deals now require secured components years ahead (Jabil consignment model), and a 3-year Lenovo master agreement with a 7-year statement of work.
• For enterprises evaluating AI clouds: a concrete case that Akamai-class distributed infrastructure is being trusted at the highest AI tier — vendor-concentration and latency arguments for agent workloads now have a flagship reference.

• What does Anthropic actually run on Akamai? "CPU workloads" could mean agent sandboxes/tool execution, retrieval, pre/post-processing, media services or something unannounced (e.g., privacy-sensitive services requiring non-hyperscaler infrastructure). Neither party said.
• Does the expansion happen? The $9B option requires Anthropic to choose it ("at mutually agreed upon terms") — watch 2027: if Claude usage compounds, 1%/tranche vesting is cheap for Akamai; if inference efficiency collapses unit demand, the option (and 3% of warrant) goes unexercised while $5.5B of capex is sunk.
• Will other clouds/labs copy the customer-warrant structure? AMD-OpenAI set the supplier-warrant-to-purchaser precedent; expect negotiation checklists to include one. Which second-tier provider (StackPath-class, regional neoclouds) lands a frontier deal next?
• Anthropic IPO arithmetic: does the warrant/relationship disclosure obligation in an eventual S-1 (material contracts, related-party-like structures) re-frame the deal publicly — including Anthropic's own accounting of a supplier equity instrument?
• Akamai's exclusivity with itself: can it keep serving hyperscaler customers and a lab that competes with them (Leighton says both; Microsoft/Google are Akamai cloud/security customers)? What do Amazon (also an Anthropic investor and a competitor's cloud) and Google make of a customer-equity supplier?
• Margin math truth: the 10-Q (filing MSA as exhibit) and 2027 quarterly disclosures will show actual recognized revenue vs the $150–300M guidance; ramp-margin pressure ("next several quarters") vs the ~30% EBIT analyst estimate needs evidence.
• Power/colo footprint: which providers/regions secure the 95–105 MW portfolio; any behind-the-meter moves; how much of the US-majority footprint gets built vs leased.
• Memory-market knock-on: does a $1.7B pre-buy by a #4-ish cloud force other buyers (or trigger supplier pushback/pricing) in 2027 contracts?

Days–weeks: Akamai's Q3 report (quarter ending Sept 30) will file the Anthropic MSA and Lenovo agreement as 10-Q exhibits (8-K commits to this) — first look at actual contract text and the warrant's booked fair value; expect analyst note cascades and further target moves (UBS had already gone $143→$148 by Sept 25 per TipRanks listing). Q4 2026: the ~$1.7B memory/component pre-buy executes through Jabil consignment; watch for supplier (Micron-class) commentary on cloud pre-purchases. 2027: revenue "begins in second half of 2027" — every guidance tick (vs $150–300M) and site-ramp margin disclosure is testable; the first warrant tranche vests on Anthropic's first Project Plan 3 payment (treasury-stock dilution math becomes real). 2028: full-ramp year → ~$1.7B run-rate proof point; expansion option ($9B/3% warrant) decision window opens in earnest. 2029–2033: contract plateau; warrant expiry Sept 18, 2033 with ~$862M cash-exercise decision for Anthropic. PREDICTION: (a) at least one other public compute supplier (neocloud or telecom-class edge provider) announces a customer-warrant-linked deal within 12 months, normalizing "supplier pays for certainty in equity"; (b) Anthropic exercises an expansion tranche (first +$3B) by end-2027, because its growth curve and IPO-window optics both reward pre-committed capacity — making the ~$20B framing real and pushing Akamai's CIS backlog past $20B YTD; (c) ramp-quarter margin pressure produces exactly one angry-earnings-cycle drawdown for AKAM in 2027 before the run-rate arrives, a repeat of this week's spike-and-fade pattern on fundamentals rather than sentiment; (d) the Anthropic silence breaks indirectly — the company's eventual S-1 risk-factor/contract disclosure becomes the first primary document stating what it runs on Akamai; (e) memory prices stay elevated into 2027, and this kind of pre-buy becomes a standard 8-K line item across the sector, drawing FTC-style commentary on circular AI finance (which the FTC chair this week already warmed to, per Reuters).

The headline number was $11.6 billion; the story is two quieter sentences from the documents. First: the contract is for CPUs — the general-purpose workhorses that agents actually run on — and a frontier lab just signed seven years of take-or-pay for them at a scale previously reserved for GPU empires. Second: the money in AI infrastructure has started flowing in both directions — Akamai gave its customer a 5% equity option, inverting the supplier-buys-the-lab pattern that critics call circular; the CFO called the warrant "very, very small," which is precisely the moment a structure becomes a trend. Watch the faded stock chart too: +22% in the dark, +3.2% by the close — the market believes the backlog and is still waiting for the ramp. Anthropic, characteristically, said nothing at all. The company that never comments on compute just bought its supplier's future stock at $111.33, and told a competitor-protected, outage-exitable, memory-hoarding cloud exactly what every enterprise buyer secretly wants: the right to walk away if the vendor is ever acquired by its rival. That's not a press release. That's a procurement manual.