Akamai Signed the Biggest Contract in Its History. The Money Isn't for GPUs.

The version of this story that led the wires is simple. Akamai signed an $11.6 billion, seven-year cloud agreement with Anthropic on 24 September, the shares jumped in after-hours trading, and Tom Leighton called it the largest contract in Akamai's history. Every one of those things is true.

Here is what the headline leaves out. The commitment is for CPU capacity, not GPU training clusters. Akamai expects no revenue from it in 2026. And the first large sum it is committing against the deal is for memory.

What Akamai actually signed

Two project plans under a master services agreement dated 5 May 2026, executed on 18 September and announced six days later. Anthropic commits to pay roughly $11.6 billion across the two plans, each with an initial seven-year term from its own service start date. The figure is subject to termination rights and conditional on Akamai meeting delivery and service-availability requirements, per the 8-K filing.

Akamai frames an option to add up to $9 billion more, which would take the total to roughly $20 billion. CFO Ed McGowan described that as an option for future business rather than a milestone-dependent extension of the existing commitment, and said any expansion would require further capital spending.

Then there is the part Akamai has not done before. It issued Anthropic a warrant for 387,051 shares of Series B non-voting convertible preferred stock, convertible into 7,741,020 common shares, or about 5% of common stock outstanding. The exercise price is $111.33 per common-equivalent share, derived from the 30-day volume-weighted average price before issuance. Roughly 2% vests on the initial $11.6 billion commitment. The remaining 3% vests in three tranches as Anthropic commits each additional $3 billion.

Leighton told Bloomberg this was the first time Akamai has agreed to a warrant as part of a cloud deal with a customer. "It's a serious step, but I think in this case it made sense to do," he said. "It helps bring the companies together."

The warrant is not free money. McGowan said its grant-date value is deducted from revenue over the life of the agreement.

The revenue arrives late. The spending does not.

Akamai expects nothing from this contract in 2026. Revenue begins in the second half of 2027, with management guiding $150 million to $300 million for that year, ramping toward an annualised run rate of about $1.7 billion by the end of 2028.

Against that, the capital plan is $5.5 billion. McGowan broke it down as roughly $1.7 billion in the fourth quarter of 2026, about $3.1 billion across 2027, and around $700 million in 2028. That is 47 cents of capital committed per dollar of contracted revenue, with most of it spent before the money starts coming back.

The Q4 figure is the one to read closely. Akamai says the $1.7 billion increase is to secure and pre-purchase critical supply chain components, "including memory." Separately, the 8-K discloses that Akamai authorised Jabil to purchase approximately $1.7 billion of memory components, with Jabil holding unused stock on consignment and repurchasing it at cost as it is drawn down.

Those two numbers match, and the filings do not say they are the same dollars. Treat them as two disclosures rather than one.

Either way, put it beside the ordinary run rate. Akamai's whole second-quarter capital expenditure was $347 million, or 32% of revenue. A single quarter's memory pre-buy is nearly five times that. For a company that spent most of its life buying servers for a content delivery network, this is a different kind of balance sheet event.

CPUs, and why that is the real story

Akamai's release is specific: Anthropic is buying CPU workload capacity. Bloomberg's Lynn Doan reported the same, noting that CPUs have seen renewed demand in data centres supporting AI services. McGowan confirmed on the call that the Anthropic deal is entirely CPU-based, while Akamai's wider portfolio mixes CPU and GPU.

The logic is not hard to follow, and Akamai has been arguing it for a year. Agentic workloads spend most of their wall-clock time on tool calls, API lookups, context retrieval and code execution, all of which run on CPUs, with the GPU sitting idle between reasoning steps. Akamai puts a number on it: CPU-side processing can account for up to 90.6% of total latency in agentic workloads. Its own State of AI Inference 2026 survey of 200 practitioners found 82% of organisations need end-to-end response times of 500 milliseconds or less, 64% need under 250 milliseconds, and half of production deployments miss their own latency targets at peak load.

Akamai's commercial answer is placement, and this contract is the first one at a scale that tests it. CPU-heavy orchestration runs near the data, GPU reasoning sits wherever it earns its keep. Akamai has roughly 4,000 edge locations inherited from the CDN business and about 20 GPU sites, with plans to move toward 100.

The honest caveat is that Akamai is a party to this argument and sells the fix. The 90.6% figure comes from a paper Akamai cites rather than publishes, and the latency survey is Akamai's own. What is not in dispute is the contract: a seven-year CPU commitment large enough that memory and power become the binding constraints rather than the server order.

Where the tension sits

McGowan says the structure is take-or-pay, which converts the headline into bankable backlog once capacity is delivered and reduces the risk of building expensive infrastructure a customer then leaves idle. For the seller, that is genuinely better than a usage-based deal.

The costs are real regardless. Depreciation starts when equipment enters service, and new colocation sites typically take 60 to 90 days to reach full revenue potential, which management flagged as temporary margin pressure during the ramp. Akamai is funding the build from $4.6 billion in cash and marketable securities reported at 30 June, a $1 billion revolving facility, and $3.5 billion in zero-coupon convertible debt raised in May. It has paused share repurchases to redirect capital into the cloud pipeline.

Then there is the circularity question. Bloomberg noted that some investors have raised concerns about deals where companies buy each other's products and invest in each other, which critics say makes real AI demand harder to measure. An equity warrant tied to future spending commitments sits squarely in that territory. TipRanks framed the after-hours reaction as two-sided: enthusiasm for the contract against concern that a large long-dated obligation plus up to 5% dilution could strain financial flexibility.

The disclosure is also thin. Akamai did not publish capacity, service pricing, a deployment timetable, or how the workloads will be distributed geographically, and the release offers no performance benchmarks, committed service levels or workload volumes. Anthropic was not quoted in the announcement.

One thing worth flagging rather than smoothing over: the share move is reported inconsistently. Reuters syndications carried 15%, 16% and 22% in different versions of the same story. SiliconANGLE said more than 20%, TradingView put it at 20.7% to $133.21, and Bloomberg reported as much as 17% to $129.60. The stock had closed the regular session down 6.78% at $110.41. Anyone quoting one precise figure is picking a source, not stating a fact.

What it means for anyone buying hardware

Akamai's own numbers say the unit of risk has moved. McGowan said the combined $14.4 billion in major multiyear cloud contracts signed this year should produce about $2.2 billion in annual recurring revenue once fully ramped, across an estimated 95 to 105 megawatts, or roughly $22 million of annual revenue per megawatt. That is a useful benchmark, and it is the arithmetic that makes memory and power procurement the gating item.

The wider pattern is that Anthropic is spreading its compute commitments across many suppliers this year: a $10 billion deal with Volta in August, a 20-year lease for about 401 megawatts at TeraWulf's Hawesville, Kentucky campus, a Micron memory and storage agreement signed in June, and, as Reuters reported, a $45 billion arrangement at Nscale's West Virginia campus. Committing roughly $4.8 billion of the $5.5 billion capital plan before the contract produces meaningful revenue is now the standard shape of a frontier lab deal.

For hardware companies, the lesson is about sequencing. When a customer commits $11.6 billion across seven years, the contract value is not what decides whether the deal works. What decides it is whether you can secure the memory, the power and the colocation on a schedule that lets depreciation and revenue move together, and whether the capital markets stay open long enough to fund the gap between the two.

That is a sourcing and cost-modelling problem wearing a procurement badge, and it is the same class of problem every hardware company in this cycle now faces.

The companies that get this right treat memory, power and colocation as the critical path rather than an afterthought to a compute order. At DMC, we work with hardware teams on exactly that: sourcing strategy, cost modelling, and production ramp planning when the commitment runs years ahead of the revenue it funds. If your roadmap has a two-year gap between spending and shipping, let's talk.