How Akamai Became Anthropic's $11.6 Billion Landlord

A Relationship Reversed in Five and a Half Months

On April 10 this year, when Anthropic launched Claude Managed Agents, Akamai’s stock fell 13% in a single day. Fastly dropped 18%, and Cloudflare fell 11%. The read was that if Anthropic hosted agents itself, CDN¹and edge computing companies would lose their business. Akamai’s stock was cut in half within a week, sliding to $64 by April 14.

Five and a half months later, on September 24, the same two companies announced the biggest deal in Akamai’s 28-year history. Anthropic committed to spending $11.6 billion on Akamai Cloud over seven years, with usage-based additions that could add up to $9 billion more—bringing the total to roughly $20 billion. Contrary to the market’s initial read that agents would push CDN companies out of the picture, it turned out that CDN companies would be the ones supplying the CPUs needed to run those agents.

The contract specifies CPUs. Akamai’s press release said the deal would support Anthropic’s growing CPU workloads, and made no mention of GPUs. It’s true that Akamai has been building out inference infrastructure by bringing in Nvidia GPUs this year, but that’s not what this particular deal is about.

I think this deal adds one more line to the list of data center players. After hyperscalers, neoclouds, colocation providers, and Bitcoin mining infrastructure, now CDN operators and mobile carriers join the list too—six categories in all. But making the list and actually making money from the business are two different things. Let’s first read the terms Akamai actually got, and then break down all six categories.

What the Contract Actually Says

Based on Akamai’s 8-K filing with the SEC, here’s what the deal looks like.

Money moves project by project. Anthropic has committed to paying roughly $11.6 billion combined across multiple project plans, and each plan runs for 7 years starting from the day service goes live. The clock starts not on the announcement date but on the day the infrastructure actually switches on. Akamai is targeting late Q2 2027 for the first service launch.

Payment comes with conditions attached. Akamai has to deliver the promised infrastructure on time and meet service availability requirements. If there’s a major outage, Anthropic can terminate that particular plan, and the master agreement includes clauses allowing either side to terminate for specific causes, including uncured breaches. The full terms are set to be attached to the Q3 10-Q, so for now we only know what’s summarized in the filing.

But once delivery is complete, Anthropic pays regardless of whether it actually uses the capacity. It’s a take-or-pay² structure. Akamai CFO Ed McGowan said that once revenue hits its stride, it stays flat for the rest of the contract term, and there’s no obligation to swap out hardware during that period either. Depreciation, he said, is scheduled to align with revenue recognition. So from Akamai’s perspective, this is close to a leased asset with a price locked in for 7 years once it’s built.

Akamai isn’t building the space itself. It’s filling the volume for this deal and other large cloud contracts by leasing data center capacity from colocation³ providers. Some of that capacity, they said, is already secured. Across all cloud infrastructure contracts signed so far, the total comes to roughly $2.2 billion in annual revenue and 95–105MW (megawatts) of power, which works out to about $22 million in annual revenue per MW across the portfolio. Akamai hasn’t disclosed the power scale specifically for the Anthropic deal, but Oppenheimer estimated it at roughly 77MW by dividing the $1.7 billion in annual revenue by that same per-MW rate.

This is where things diverge from the picture the name “CDN” tends to conjure. It’s not servers sprinkled thinly across 4,300 CDN points of presence around the world — it’s CPU servers packed densely into tens of megawatts of space leased from colocation providers. Neither company has disclosed exactly what workloads Anthropic plans to run on that CPU capacity.

Who Builds Data Centers? - Haebom’s Archive - Kwangseob Ahn ex)haebomData centers underpin every digital service we use day to day — cloud services, streaming platforms, online shopping. But how this vast, complex facility is put together, and who builds it, isn’t widely understood. We tend to speak vaguely of “cloud infrastructure” or “data centers,” but this industry is far broader and more diverse than it sounds, combiningslashpage.com

If you want more on data centers, I wrote a piece on this about 2 years ago — worth a look.

The Order the Money Moves

The market’s real concern is sequencing. According to Akamai’s investor presentation, the capital expenditure (CAPEX) built into this contract runs to about $5.5 billion. Of that, $1.7 billion goes out in Q4 of this year, $3.1 billion in 2027, and $700 million in 2028. Most of the Q4 $1.7 billion is money spent pre-purchasing components like memory. Akamai also announced that it will source roughly $1.7 billion worth of memory through Lenovo and Jabil.

Revenue arrives late. This year it’s zero. It starts showing up from the second half of 2027, with full-year 2027 revenue expected at $150–300 million. Annual revenue only reaches the full run-rate of roughly $1.7 billion by the end of 2028. Looking at the end of 2027, Akamai will have spent a cumulative $4.8 billion while earning, at most, $300 million.

Seeing how these numbers stack up against Akamai’s own size explains the market’s reaction. Before the announcement, Akamai’s market cap was about $15.9 billion, its 2026 revenue guidance was $4.4–4.55 billion, and its 2025 free cash flow was around $636 million. The $5.5 billion in CAPEX is more than a third of the market cap, bigger than a full year of revenue, and would take more than 8 years of free cash flow to cover. Akamai had already issued $3 billion in convertible bonds⁴ across two tranches back in May. This latest announcement didn’t lay out any additional fundraising plans, but under the current structure, the picture is one of building with debt and cash flow now, and recouping it over 7 years.

Capital efficiency is also low. Dividing the fully-ramped annual revenue of $1.7 billion by the $5.5 billion CAPEX yields a ratio of about 0.31x — well below the ratio Akamai has historically maintained on cloud contracts. By Oppenheimer’s math, CAPEX per MW runs about $70 million, of which $22 million is memory pre-purchases. That means memory costs account for nearly a third of total CAPEX — a detail Korean readers, in particular, might read through a different lens.

The stock price tracked this math exactly. The day after the announcement, on the 25th, shares jumped over 20% before the open, rose as much as 16.4% intraday to $128.46, then closed at $113.87, up 3.1% — 11% below the day’s high. Analyst reactions were split. BofA ($185), Guggenheim ($225), and Evercore ($175) raised or held their price targets, while JPMorgan ($167), UBS ($148), and Citi ($122) stayed neutral. Evercore argued that CPU workloads generate higher revenue and cash gross margin per MW than GPU workloads, and that operating margins would move to the upper end of the previous framework (low-to-mid-20s% to low-30s%). Citi, on the other hand, cut its price target, citing the CAPEX burden.

One more simple calculation. The $11.6 billion in contracted revenue over 7 years is about 2.1x the $5.5 billion CAPEX. That’s before subtracting electricity, rent, and operating costs, so it’s not profit. Even so, this contract shifts a large share of demand risk onto Anthropic. Once delivery is complete, the money keeps coming in for 7 years. What’s left for Akamai is the risk of actually delivering the infrastructure on the promised date at the promised quality. A company whose contract size rivals its own market cap taking on a deal with delivery conditions attached also means it has shouldered execution risk at the scale of its entire company.

Why CPU, and Why Akamai

Most of the computing deals Anthropic has signed this year are measured in gigawatts. With AWS: over $100 billion across 10 years, up to 5GW of Trainium. With Google and Broadcom: roughly 3.5GW of TPUs starting in 2027. With neocloud Enscale: about $45 billion over 6 years for 460MW in West Virginia. Fluidstack at $50 billion, Lambda at $35 billion, CoreWeave, and even SpaceX’s Colossus. Depending on how you tally it, this portfolio runs anywhere from $275 billion to over $500 billion. Next to all that, the Akamai deal — $11.6 billion, an estimated 77MW — looks like a small line purely by size.

But it’s a different kind of deal. The contracts above are accelerator deals for training and running inference on models. The Akamai deal is a general-purpose compute contract for when those models actually go to work. As I covered in Issue 217, agents think on GPUs and work on CPUs. The moment of planning is brief; the rest of the time — opening browsers, running code, reading files — plays out on CPUs. Managed Agents, the product Anthropic launched in April, is exactly that kind of thing. Back in April, the market read it as a competitor to Akamai — but running it requires someone to operate CPU servers at tens of megawatts of scale. Operating CPU servers scattered across the globe is precisely what Akamai has been doing for nearly 30 years.

There are two clues to why Anthropic chose Akamai over a hyperscaler. One is speed. Last week, CNBC reported that Anthropic and OpenAI are hunting for smaller data center deals in the 20-30MW range. The reasoning: gigawatt-scale campuses won’t come online until 2027-2028, while smaller blocks can be put to use faster. The 77MW that Akamai is bringing online through leased colocation in late Q2 2027 fits that math exactly. The other is unit economics. Evercore noted that CPU workloads generate more revenue per megawatt than GPU workloads. Fill the same power envelope with GPU servers, and most of the cost sits in the GPUs themselves, leaving a thin margin for the cloud operator; CPU servers carry relatively lower component costs, so more of the value stays with the operator.

Let me note what remains unconfirmed here. Why Anthropic chose Akamai over CPU instances from its own hyperscaler partners, and whether Akamai’s distributed points of presence are actually being put to use — or whether Anthropic is simply drawing on a CPU pool packed into colocation facilities — hasn’t been disclosed. Whether Akamai’s long-standing asset, its 4,300 points of presence, is what’s actually being valued in this deal, or whether Anthropic simply picked a company that could stand up CPU servers quickly, is something we’ll only know once future disclosures and actual deployments come in.

The Supplier Gave the Customer Equity

The contract contains a clause Akamai has never attached to a cloud deal before: Akamai issued Anthropic a warrant⁵. It’s the right to buy non-voting convertible preferred stock—convertible into roughly 7.74 million common shares—at $111.33 per share. At most, that comes to about 5% of Akamai’s outstanding shares, and the strike price sits near the closing price from the day before the announcement.

That 5% doesn’t vest all at once. About 2% (roughly 3.1 million shares) is locked in once Anthropic makes its first payment, and the remaining 3% vests in 1% increments each time Anthropic commits an additional $3 billion. If Anthropic exercises the full $9 billion option, the whole 5% fills up. JPMorgan described the path toward exercising this option as the “path of least resistance.”

In the AI industry, circular deals usually run the other way. Amazon and Google invest in Anthropic, and Anthropic uses that money to buy AWS and Google Cloud. Microsoft and Nvidia agreed late last year to put up to $15 billion into Anthropic, and in July of this year AMD invested up to $5 billion in Anthropic while locking in a 2GW MI450 deployment. It’s a pattern where the supplier becomes the customer’s shareholder. This time it went the other way—the supplier gave the customer equity. A similar case is the warrant AMD gave OpenAI last year: a structure releasing up to 160 million AMD shares, about 10%, tied to OpenAI’s AMD-hardware deployment milestones and AMD’s share price was the arrangement.

Lay the two directions side by side and a rule emerges: equity doesn’t flow to whoever puts up capital first—it flows to whoever makes the commitment first. In AMD-OpenAI, OpenAI committed to hardware purchases; in Akamai-Anthropic, Anthropic committed to seven years of payments. Thanks to that commitment, the supplier can borrow to build, so it hands over equity as the price of the promise. Akamai was able to issue $3 billion in convertible bonds in May precisely because of the $1.8 billion contract signed that same month. From shareholders’ perspective, 5% dilution is a cost—but without that 5%, the $11.6 billion commitment likely wouldn’t exist at all.

How differently a promise gets priced shows up, in hard numbers, in another company’s disclosures. This is a case I covered in a ZDNet column last week: in the fiscal-year results that Nasdaq-listed GPU cloud operator IREN disclosed on August 27, the $3.6 billion in GPU financing tied to its Microsoft contract was raised at an investment-grade rate of 6.0% annually. The $2.4 billion in financing used for deployments to non-investment-grade customers, by contrast, was funded by investors advised by Blue Owl and PIMCO at a fixed 9.0% annual rate. Same company, same equipment—yet the interest rate diverges by 3 percentage points depending on who the tenant is.

Akamai took a different path. Rather than project financing backed by Anthropic’s credit, it issued zero-coupon convertible bonds in its own name. The interest rate is near zero, but in exchange it left the door open to dilution through conversion—and opened it again with the warrant. If IREN lowered the cost of money using its tenant’s credit, Akamai lowered it using its own stock. And if Anthropic’s payment commitment ever falls through, that debt lands squarely on Akamai’s shareholders.

Data-center operators now come in six flavors

Let’s return to the list. Companies supplying compute to AI data centers can be broken down like this:

Category	What they already had	What they sell	Recent example
Hyperscaler	Its own cloud, capital, its own chips	Cloud services and dedicated clusters	AWS–Anthropic deal worth over $100 billion over 10 years, Trainium 5GW
Neocloud	GPU-sourcing capability and fast build-out	GPU time, dedicated campuses	Enscale–Anthropic 460MW deal worth $45 billion, CoreWeave
Colocation	Buildings, power, cooling	Space and electricity	The leased space Akamai will use to fill this order
Bitcoin mining infrastructure	Large-scale power contracts and sites	Power-equipped buildings, sometimes GPUs too	IREN’s Microsoft contract, $3.6 billion GPU financing
CDN operator	Distributed server operations capability, network	CPU compute	Akamai–Anthropic deal worth $11.6 billion over 7 years
Telecom carrier	Central offices, power, fiber-optic cable, enterprise customers	Data-center space, circuits	SK Telecom’s 15GW plan, Verizon–Google $1 billion dark fiber deal

Each row arrived by a different route. Hyperscalers are companies that originally built data centers to run their own services. Neoclouds[^6] started out by securing GPUs first and then leasing them out. Colocation is fundamentally a real-estate business — selling buildings and electricity regardless of what the servers inside are doing. Bitcoin mining infrastructure operators held power contracts and sites, and when mining profitability declined, they redirected those assets toward AI. CDN operators used to run CPU servers scattered around the world, and telecom carriers hold central offices, fiber-optic cable, and sites with power already brought in.

Reading this table as a set of layers is misleading. These six categories aren’t different layers — they’re different origins. Just look at this one deal: Akamai, a CDN operator, is leasing space from a colocation operator. Neoclouds are also frequently tenants of colocation providers, and bitcoin mining infrastructure operators lease buildings to neoclouds. Each player walks in with one asset accumulated from its original business, and borrows whatever else it lacks from a different type of operator.

That borrowing comes at a price. The risk isn’t symmetric between the borrower and the lender. Colocation and mining infrastructure operators only sell buildings and power, so it doesn’t matter to them what happens to server prices. But if the tenant fails, they’re left holding an empty building. Neoclouds and CDN operators, by contrast, buy the servers themselves — that 0.31x figure for Akamai is exactly that cost. In exchange, they collect fees for operating the servers, so their revenue per MW is higher. The further up you go, the more capital it takes; the further down, the less capital but also the less money you make. Reports that Anthropic and OpenAI are hunting for 20–30MW contracts mean opportunity is widening for the bottom four rows of this table — the operators who already own the power and the buildings.

Korea’s Sixth Line

There’s a Korean company on the sixth line. In an SEC 6-K filing this year, SK Telecom disclosed that it’s reviewing a mid-to-long-term plan to build a total of 15GW of AI data centers in phases. The plan calls for opening 5GW sequentially starting in 2029, then adding another 10GW from 2035. In July, it established a subsidiary, SK Hyper, to handle site acquisition, substations, construction, and customer acquisition.

A new AI DC business development company launches to build an ‘Asia AI Infrastructure Hub’ - SK Telecom NewsroomSK Telecom is taking its first step toward building an ‘Asia AI Infrastructure Hub.’ At a board meeting held on the 23rd, SKT resolved to establish a new subsidiary dedicated to AI DC (data center) business development, ‘SK Hyper Co., Ltd. (SK Hyper),’ and to invest ₩750 billion (~$540M) by 2030 to lay the groundwork for the business. Under rapidly shifting AI competitive conditions, SK Hyper will enable fast decision-making andnews.sktelecom.com

Let me place this number next to the Akamai contract. Akamai is building roughly 77MW for $11.6 billion. 5GW is 65 times that. Omdia has warned about the AI infrastructure plans of Verizon, AT&T, and SK Telecom: the headline numbers are ceilings, and revenue will lag capacity growth by years. The point is that hyperscalers have a habit of writing the final cap into the contract while only buying a fraction of it at the outset.

I’ve been traveling to various regions lately, advising on, designing, and reviewing AI data center projects. When I get a review request, I don’t ask about the site or the GPU count first. I look first at whether there’s a confirmed tenant, and whether the lending syndicate is moving on that tenant’s credit. If there’s no answer to that, the rest of the paperwork isn’t worth reading yet.

Without a tenant, power approval doesn’t come through; without power approval, financing doesn’t close; without financing, you can’t decide what to fill the facility with. The reason Korea’s National AI Computing Center bid failed twice in 2025 wasn’t the GPU count or the budget — it was the equity structure: 51% public ownership plus a buyback option. Only in the third round, after the equity and buyback terms were revised, did the Samsung SDS consortium submit a solo bid.

By that standard, the Akamai deal is a case study in getting the sequence right. The tenant was confirmed first ($1.8 billion in May), that contract was used to open financing ($3 billion convertible bond in May), and only then was the equipment decided ($11.6 billion in September, plus $1.7 billion in advance memory purchases). If an anchor customer attaches to the 15GW plan, the same kind of terms will follow: who bears demand risk, who bears delivery risk, and which direction equity flows. Until those terms are disclosed, the GW figure is about as informative as a lot’s square footage.

One more thing on memory: of Akamai’s $5.5 billion in capex, $1.7 billion is advance memory purchases. Because these are CPU servers, memory fills the slot that would otherwise go to GPU spending. Growth in CPU infrastructure for agentic workloads means growth in demand for ordinary server DRAM, not HBM — and that’s a different kind of memory order than what GPU servers require.

Oswarld’s Lens

Last week, I wrote in a ZDNet column that AI data centers start out as real estate development and end up as power businesses and equipment finance. Buildings last 30–40 years, but the GPUs inside them turn over every 2–3 years, and in Korea, many deals are full-service leases where the landlord provides the building plus power and cooling infrastructure — meaning the landlord absorbs that depreciation. It’s a pattern I’ve seen repeat across every region I’ve visited.

The True Nature of AI Data Centers… A Complicated Four-Variable FunctionLately I’ve been traveling to various regions to advise on, design, and review AI data center projects. The type I meet most often is the one who treats this business too lightly — the math being: if there’s a site and power nearby, you just build. Well, that’s partly true. Which is why I wrote this piece. AI data centers in Korea right now…zdnet.co.kr

Reading the Akamai contract again through that lens, one clause jumps out: there’s no obligation to replace hardware during the contract term, and depreciation is matched to revenue recognition. If this were GPUs, generations would turn over two or three times within a 7-year contract, making this clause nearly impossible to sustain. Neocloud contracts running 5–6 years typically build in a default “mutual agreement upon equipment generation change” clause as a matter of course, and for a fixed 7-year term with no replacement obligation to hold up, you’d have had to lay the GPU depreciation curve directly onto the contract itself. At today’s market prices, that’s simply not possible. This is a condition that only works because it’s CPUs.

In my advisory work, when I look at service terms for CDN operators or telecom carriers, CPU-based contracts almost always assume “no changes during the operating period.” The “no replacement obligation” in the Akamai-Anthropic deal is an extension of that convention, and that’s precisely what makes a 7-year term viable. While the market was busy being disappointed by the low capital efficiency, I was looking harder at this clause. Akamai sold CPUs instead of GPUs partly because CPUs were the only asset it had — but that also let it push the most dangerous problem in the data center business, equipment that depreciates every 3 years, off onto the contract itself.

In that column, I called Nvidia a “furniture company.” The point was to warn against buying furniture before you have a tenant for the house — but Nvidia has now gone beyond selling furniture, guaranteeing up to $105 billion of OpenAI’s rent and power costs at the 4.25GW Ohio campus. There’s no such guarantor in the Akamai deal. Anthropic took equity instead of offering a guarantee, and Akamai borrowed against its own stock as collateral. So the risk in this contract concentrates on two questions: can Anthropic keep paying for 7 years, and can Akamai turn the lights on at the promised quality by late Q2 2027. These are the same four things I check first when reviewing contracts across six different types of operators: who bears the demand risk, who bears the delivery risk, who bears the equipment depreciation, and which direction the equity flows.

Reading April and September together makes this deal clearer. In April, the market saw agents as stealing work from infrastructure companies. In September, agents became infrastructure companies’ biggest customers. Both reactions came from the same underlying fact: agents run on top of CPUs, and someone has to operate those CPUs. April’s reading wasn’t wrong — some of the edge computing services Akamai used to sell are now provided directly by Anthropic. What September added was the information that Anthropic chose not to operate the servers those services run on itself.

Closing

Akamai secured $11.6 billion in confirmed revenue over seven years, and in exchange, it agreed to spend $4.8 billion upfront by the end of 2027 and open up to 5% equity to its customer. Data center operators can be sorted into six categories, from hyperscalers to telecom carriers, and this was the deal terms for the first major contract signed by the category that arrived last: CDN providers. The invitation still stands, but the tenant is holding the admission ticket—and now that tenant is asking for equity, too.

---

💬 Of the six categories, which operator do you think will land the most AI infrastructure deals going forward? Drop a comment with one line explaining why.

📨 If you know a colleague who still thinks of AI infrastructure investment purely as a race to secure GPUs, share this piece with them.

---

References & Further Reading

• Akamai, press release, 2026.9.24. Original source for contract value, expansion options, and warrant terms.
• Akamai, Form 8-K Exhibit 99.1, SEC filing, 2026.9.24.
• TechCrunch, Anthropic to pay Akamai $11.6 billion over seven years in cloud deal, 2026.9.25. Delivery and availability terms, termination clauses, warrant vesting stages, comparison with AMD and OpenAI warrants.
• Mugglehead, Akamai signs US$11.6B Anthropic cloud deal with US$5.5B buildout, 2026.9. Cites the 8-K’s project plan, 7-year start date, termination conditions, and quarterly capex.
• BankInfoSecurity, 7-Year, $11.6B Anthropic Deal Drives Akamai Cloud Buildout, 2026.9.25. Take-or-pay structure, colocation leasing, 95-105MW and $22 million per MW.
• CNBC, Shares of Akamai surge after deal with Anthropic. What Wall Street is saying, 2026.9.25. Intraday high of $128.46, Oppenheimer’s estimate of 77MW at $70 million per MW, Evercore’s assessment of CPU margins.
• TECHi, Akamai’s $11.6B Anthropic Deal Comes With $5.5B Capex, 2026.9. Closing price of $113.87 on the 25th and intraday movement.
• Investing.com, Piper Sandler raises Akamai stock price target on Anthropic deal, 2026.9.25. Analyst target price changes, pre-announcement market cap.
• IFR, Akamai Technologies funds AI cloud capex with US$3bn CB, 2026.5. The $3 billion convertible bond issuance in May and the structure of existing convertible bonds.
• Boston Globe, Akamai Technologies’ stock price surges on big AI deal, 2026.5.13. The stock’s halving and $64 low after the April Managed Agents launch, and the $1.8 billion deal in May.
• Seeking Alpha, Fastly, along with Akamai and Cloudflare, tumbles after Anthropic launches Managed Agents, 2026.4.10. The drop across all three companies on April 10.
• CNBC, Anthropic and OpenAI hunt for smaller data center deals, 2026.9.18. The search for smaller 20-30MW deals, and Enscale’s 460MW.
• Anthropic, Anthropic expands partnership with Google and Broadcom, 2026.4.6.
• Value Add Pulse, Anthropic’s AI compute deals top $275 billion in 2026, 2026.9. A tally of deal sizes with AWS, Fluidstack, Enscale, SpaceX, and Lambda.
• Wikipedia, AI build-out financing. Microsoft, Nvidia, and AMD’s investments in Anthropic, and the warrant terms in the AMD-OpenAI deal.
• SK Telecom, Form 6-K, SEC filing, 2026. The 15GW AI data center plan.
• RCR Wireless, Building the AI backbone, waiting on the AI billions, 2026.7.31. Omdia’s warning on telco AI infrastructure revenue, Verizon and Google dark fiber.
• Kwangseob Ahn, ZDNet Korea, [Kwangseob Ahn’s AI Synthesis] The Essence of AI Data Centers… A Complicated Four-Variable Function, 2026.9.19. Tenant credit and financing costs, Iren’s 6.0% and 9.0% financing, Nvidia’s $105 billion guarantee, the failed bid for the National AI Computing Center, and the mismatch between 40-year buildings and 3-year equipment.
• INLEVEL9 Letter Issue No. 217, Agents Think in GPUs, Work in CPUs. Covers why CPU demand is rising in the age of agents.
• INLEVEL9 Letter Issue No. 233, AI Servers Earn 44% Margins, So Why Does Cloud Only Earn 11%?. A story about how profitability differs depending on where the same AI infrastructure sits.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. CDN (Content Delivery Network): A network of servers placed across multiple global locations that delivers copies of videos, files, and web pages from a point close to the user. Akamai has stated that it operates over 4,300 such locations. ↩

2. Take-or-Pay: A contract requiring the buyer to pay for a committed volume even if they don’t actually use it. This reduces the supplier’s exposure to demand fluctuations, but often comes with delivery obligations in return. ↩

3. Colocation: A leasing arrangement in which a data center operator provides the building, power, cooling, and network, while the customer places its own servers inside. ↩

4. Convertible Bond: A bond that can be converted into stock under set conditions. It carries a lower interest rate in exchange for the possibility that, if the share price rises, it converts into equity and dilutes existing shareholders. ↩

5. Warrant: The right to buy a company’s stock at a set price. In this deal, the exercisable volume increases based on how much Anthropic spends. ↩