Akamai Plans $1.7 Billion in Spending Before Revenue From Its New Anthropic Deal

Akamai's September 24 disclosure of a seven-year, $11.6 billion Anthropic cloud commitment is an AI infrastructure story with a financing problem attached. The company plans about $1.7 billion of capital expenditure in the fourth quarter of 2026, before any revenue from this commitment. The disclosed workloads are CPU-based; describing the transaction simply as another GPU purchase would miss that distinction. The project plans were signed on September 18, so this is a fresh disclosure rather than a September 24 signing. For ZharfAI, the central question is how the supplier funds the interval between acquiring equipment and earning service revenue.

Management's presentation makes the timing unusually explicit. It estimates 2027 capital expenditure of $3.1 billion against revenue of $150 million to $300 million, with service starting in late second-quarter 2027. A further $700 million of capital expenditure is planned for 2028; the full revenue run rate is expected by that year's end, followed by roughly $1.7 billion annually. Estimated total capital expenditure is $5.5 billion. These are management projections, not completed results. Akamai has not increased its overall 2026 revenue guidance because of this announcement. The size of the commitment should not obscure the slow initial revenue ramp.

Subtracting $5.5 billion from $11.6 billion does not produce contract profit. That arithmetic leaves out operating expenses, financing, payment timing and the difference between capital spending and an accounting-period expense. Dividing the commitment evenly by seven also says little about the first year's financing needs when service capacity arrives in stages. This distinction is the practical value of the disclosure: future contracted sales and today's equipment invoices can both be substantial without balancing each other in the same period. The order establishes an important commercial relationship; it does not, by itself, explain the source or cost of every dollar needed to deliver it.

There is a separate equity component. Anthropic received a warrant for non-voting preferred shares at a common-share-equivalent exercise price of $111.33. Its potential size is roughly 5% of Akamai's currently outstanding common stock, but not all of it is available immediately: approximately 2% relates to the initial commitment, with the remainder dependent on expanded purchasing commitments. The additional $9 billion opportunity is not an already completed purchase. Akamai also says the initial warrant's fair value reduces revenue and is already reflected in the $11.6 billion figure. Subtracting it again from that disclosed amount would double-count the adjustment.

For existing shareholders, issuing a purchase right is different from paying cash today, but it is not economically irrelevant. If the right becomes exercisable and is used, existing owners' economic participation can be diluted, while the customer gains exposure to the supplier's upside. This explains the structure, not whether the deal is fairly priced or where the shares will trade. Evaluating the outcome requires both service economics and the warrant's effect. A larger sales commitment alone cannot establish how much each existing share benefits. Nor should a conditional equity right be described as an immediate transfer of the entire potential stake.

The 8-K identifies a tangible procurement step: Akamai authorized Jabil to acquire approximately $1.7 billion of memory components, with supplier invoices paid when Jabil receives them; the components are held on consignment pending use. Ordering equipment and delivering a revenue-producing service are therefore visibly different events. The same filing subjects the customer's commitment to delivery, service-availability requirements and termination rights. It would be misleading to call the whole amount cash already collected or an unconditional receipt. The summary also does not establish all customer-prepayment arrangements or every contractual provision. Those omissions matter when trying to reconstruct the complete financing schedule.

Cipher's September 25 Barber Lake disclosure provides a separate comparison. Contracted lease duration rises from ten to twenty years and contracted revenue from $3.8 billion to over $9 billion. Hall deliveries are scheduled from fourth-quarter 2026 through first-quarter 2027, with rent commencing per delivery. Cipher bears the first $359.3 million above the original budget; the tenant reimburses half of further overruns over twenty years through additional rent carrying a contracted return. The release does not name the AI lab behind the new commitment. These terms belong to Cipher's project, not to Akamai's agreement, and the distinction should remain explicit.

A longer lease can improve visibility over a building's useful commercial life without supplying all the cash needed to construct it now. Under Cipher's disclosed framework, reimbursement of part of an overrun occurs over years, rather than necessarily arriving when equipment is purchased. The useful comparison with Akamai is the timing question, not an assumption that the businesses are identical. A cloud-service supplier and a data-center landlord sell different things. Adding their headline contract values without accounting for service type and duration does not create a meaningful measure of industry profit or end-user demand. Both announcements require a delivery schedule alongside the sales number.

Power has its own milestones. On September 25, developer Gray Oak reported that its South Texas campus had advanced through ERCOT's Batch Zero large-load interconnection study process. It targets first power in 2028 and utility-grid power in 2029. That is a developer's account of progress, not an announcement that the facility is energized. Land readiness, a connection-study milestone and electricity delivery are different states. Nothing in this release establishes that Akamai depends on this particular site. The relevant comparison is the distance between an administrative or engineering milestone and an asset that can actually provide the contracted service.

The macroeconomic evidence needs an equally clear boundary. In its September 25 release, the US Census Bureau put seasonally adjusted August durable-goods new orders at $338.6 billion, virtually unchanged month over month; excluding transportation equipment, orders increased 0.3%. These are dollar values, not price-adjusted volumes, and the release does not isolate AI demand. August also precedes the September contract disclosures discussed here. A flat aggregate neither disproves an AI infrastructure expansion nor turns a large cloud commitment into evidence of a broad manufacturing boom. The data supply context about industrial demand, not a measurement of these projects' returns.

For an Iranian team budgeting infrastructure, the transferable lesson concerns payment and acceptance, not copying the scale of these purchases. In a hypothetical smaller project, paying for servers before power and networking are ready creates an asset but not necessarily sellable capacity. Procurement should specify the usable output being accepted, when service charges begin and who pays for delay or changed requirements. This is ZharfAI's budgeting interpretation, not a promise of access to a foreign service or a claim that Anthropic's terms can be replicated locally. Customer demand and the ability to deliver the promised capacity remain separate questions even when the same budget discussion covers both.

Next, follow Akamai's memory deliveries, fuller contract disclosures in its quarterly report, the planned 2027 service start and any revision to estimated capital expenditure. At Barber Lake, each hall's delivery and rent commencement will provide stronger evidence than repetition of a twenty-year revenue total. For Gray Oak, subsequent connection and power-delivery milestones matter. None of the management schedules is a guarantee. The conclusion is not that infrastructure investment is necessarily excessive or insufficient. It is that a multibillion-dollar order does not remove the supplier's responsibility to finance construction, meet delivery conditions and turn purchased components into a working service.