Akamai Stock: Why AI Traffic Makes It A Buy

I kept refreshing a quote page last week and feeling slightly ridiculous about it. A stock I had mentally filed under “steady internet plumbing” jumped, then quietly handed most of that jump back, and the commentary around it sounded louder than the price action. That mismatch is usually where the interesting work starts. Not in the headline pop. In the part where the crowd gets bored and wanders off before the underlying deal has even been digested.

The deal in question is not small. A leading artificial-intelligence lab committed roughly $11.6 billion over seven years for added computing capacity with Akamai. Around the same time, a veteran market commentator looked at three companies that sit in the path of exploding web traffic and said only one was cheap enough to buy right now. Akamai. Not the name that has already doubled and then some. Not the one trading at a multiple that assumes almost nothing can go wrong. The quieter one.

Maybe that call is early. Maybe it is simply arithmetic. I have found that when AI stories get this crowded, the useful question is rarely “is the theme real?” The theme is real. The useful question is who is still being paid like the theme has not arrived.

The Quiet Tollbooths Of The AI Internet

One side effect of the AI rush does not show up in model-release videos. It shows up in request logs. Agents do not browse the way people do. They do not get tired, they do not close a tab after three minutes, and they do not politely wait their turn. They fetch, retry, scrape, summarize, and call tools in loops. Somebody has to keep that flood from turning ordinary websites into spinning wheels.

That somebody, in practice, is a small club of internet infrastructure firms. Cloudflare, Akamai, and Fastly run networks of servers that cache content, absorb junk traffic, terminate connections closer to the user, and generally act as traffic cops for the public web. Their job used to be described, a bit dismissively, as content delivery. Faster video. Fewer outages on shopping days. Useful, yes. Glamorous, no.

Glamour is the wrong scorecard now. If agents really do generate far more web activity than the humans they are supposed to help, the firms that sit between those agents and the open internet stop being background utilities. They become capacity businesses. And capacity businesses get re-rated when a large customer writes a multi-year check.

What These Networks Actually Sell

It helps to strip the jargon. A content delivery network places copies of files nearer to the person, or the bot, asking for them. A security layer in front of a site decides which requests look legitimate. An edge platform runs a slice of logic close to the user so a round trip to a distant central cloud is not required for every click. None of that is science fiction. It is routing, caching, and filtering, sold on a contract.

The newer pitch is inference, not training. Training a large model wants enormous, tightly packed clusters and a lot of power in a few places. Using a model, over and over, for ordinary tasks, wants something else: lower latency, more locations, and the ability to absorb bursts. That is closer to the map these traffic companies already own. Akamai has been talking about a network of smaller data centers that can process AI workloads nearer to users. In plain English, move some of the thinking out of the mega-hall and toward the edge.

I am not romantic about that shift. Edge is a word vendors love because it sounds inevitable. Sometimes it is just a more expensive way to do a job a central region already does well. Still, latency is not a marketing slogan when an agent is chaining ten calls to answer one question. Every extra hop is felt. The companies that already have feet in hundreds of cities start that race with a map, not a press release.

Why Agents Change The Traffic Math

Human traffic has a ceiling that product people understand. There are only so many waking hours, only so many thumbs. Agent traffic does not share that ceiling. A single workflow can fan out into dozens of page fetches, API calls, and retries. Multiply that by every company bolting an assistant onto its product, and the curve stops looking like “more users.” It starts looking like “more work per user.”

Earlier this year, Cloudflare’s chief executive told that same market host that traffic from AI agents had already surpassed human traffic on the company’s network. Sit with that for a second. Not “might surpass.” Had already surpassed, on one of the largest networks in front of the public web. Whether you love the stock or not, that is a data point, not a slogan.

The internet’s next congestion problem may not be a viral video. It may be a million polite robots asking for the same page at once.

Perhaps the most interesting aspect is how unevenly the market has priced that congestion. Year to date, Fastly shares have surged about 159 percent. Cloudflare has climbed roughly 78 percent. Akamai has gained about 22 percent. The broad market, over the same stretch, is up somewhere near 12 percent. All three infrastructure names beat the index. Only one of them still looks like a normal business on an earnings multiple.

Three Names, Three Very Different Prices

Price is not the same thing as value, and I get tired of people pretending a low multiple is automatically a gift. Cheap can mean ignored. It can also mean impaired. The job is to tell those apart before the multiple does the telling for you.

On this year’s expected earnings, the spread is almost comical. Cloudflare changes hands at roughly 279 times. Fastly sits nearer 50 times, with a history of sharp swings that can erase a thesis in a week. Akamai trades at less than 16 times. A commentator who likes the whole theme still said he would wait for a pullback on the expensive names, and that Akamai was the only one cheap enough to pound the table on. His line, roughly: you are basically getting the latest partnership for free.

I would not put it quite that cleanly. Partnerships are not free, capacity has a cost, and seven-year commitments can be renegotiated in ways retail investors never see. But the direction of the argument is hard to wave away. One stock has already been paid for a future that may arrive. Another has been paid a little, and still carries a habit of disappointing. The third has been paid almost like a mature delivery network, while signing a capacity deal that belongs to the new story.

Numbers like these move. Multiples are snapshots, not commandments. Treat the table as a map of sentiment on a Thursday in early October, not as a promise that next Thursday will look identical. The shape, though, is the point. The market has already voted with its feet on two of these names. It has been much more grudging with the third.

The Capacity Commitment That Reset The Conversation

Last week Anthropic made an $11.6 billion commitment to Akamai for added computing capacity over the next seven years. Read that slowly. Not a pilot. Not a logo on a slide. A multi-year capacity commitment measured in eleven figures, tied to the part of the AI cycle that is supposed to matter from here: using models at scale, not just training them in a handful of halls.

The stock did what stocks do. It jumped. Then it surrendered most of those gains. That giveback is the opening the commentator pointed at. Initial reaction, second thought, drift. If you have watched enough deal announcements, you know the pattern. Fast money prices the headline. Slower money asks whether the capacity can actually be built, whether margins survive the build, and whether the customer concentration is a feature or a risk.

Those are fair questions. An $11.6 billion figure over seven years is about $1.66 billion a year if you spread it evenly, which real deployments never do. Akamai’s existing revenue base is larger than a niche CDN, so this is material without being the entire company overnight. It is also a vote from a lab that can choose where to put compute. Labs do not scatter eleven-figure commitments for the scenery.

A commitment that size is not a press release. It is a claim on power, racks, and the people who know how to keep both online.

Market desk note, paraphrased

In my experience, the market under-reacts to infrastructure deals when the buyer cannot picture the product. Everyone can picture a chatbot. Fewer people can picture a smaller data center in a secondary city taking inference load off a congested region. The inability to picture it is not the same as the absence of it. It just means the re-rating, if it comes, arrives in chapters rather than in one gap-up.

Valuation Is The Actual Separator

Theme investors hate this part. They want the story, and the story is the same for all three: more bots, more requests, more need for someone to keep the pipes clear. Valuation investors live for this part, because the story being the same is exactly why price matters. You are not choosing whether AI traffic is real. You are choosing what you pay for a claim on it.

Under 16 times this year’s expected earnings is not a deep-value graveyard multiple. It is a grown-up multiple for a company with real cash generation and a legacy that the market still half-remembers as “the CDN from the early web.” That memory is the discount. Legacy is a tax until a new contract forces people to update the file.

Contrast that with 279 times. I can admire a business and still refuse to pay a price that requires years of flawless execution just to grow into the quote. The commentator likes Cloudflare. He also said wait. That is not a contradiction. It is sequencing. A wonderful asset at a careless price becomes a mediocre investment. A decent asset at a careless price becomes a lesson.

Fastly sits in the awkward middle. Fifty times is not 279, so the sticker shock is milder, but the operating history has been bumpier, and the repositioning toward an edge cloud platform is still a promise the market has learned to doubt and then suddenly believe again. Sharp swings are not a personality quirk. They are what you get when the shareholder base is mostly people trading the narrative, not people underwriting the cash flows.

• Cloudflare: strong traffic evidence, extreme multiple, better bought on a real pullback than on enthusiasm.
• Fastly: edge-cloud repositioning, mid-pack multiple, volatility that punishes impatience.
• Akamai: slower year-to-date move, sub-16 multiple, fresh multi-year capacity commitment.

If you only remember one comparison, remember that list. The theme does not pick the stock. The price does.

Smaller Halls, Closer To The User

Akamai’s infrastructure pitch is specific enough to test. Build a network of smaller data centers. Process AI workloads closer to users. Cut latency as the industry moves from training models to using them. That is a different capital plan from “rent more GPUs in the same three regions everyone else is fighting over.”

There is a logic to it. Inference that has to feel instant, or close to it, hates distance. A support agent drafting a reply, a shopping assistant checking inventory, a coding tool fetching documentation: each of those can tolerate a little delay. None of them loves a cross-country round trip on every step. Putting compute where the old delivery network already has presence is a reuse of geography. Geography, in this business, took decades to assemble.

Will every workload want that? No. Heavy reasoning may stay in big centralized clusters where the hardware is densest and the power contracts are largest. Light, repetitive, latency-sensitive calls are the slice that fits the edge story. You do not need the entire AI economy to migrate. You need a fat enough slice to fill the halls you actually build.

I keep coming back to a plain analogy. Think of training as the factory that designs the engine. Think of inference as the service stations that have to start that engine thousands of times a day, close to where the drivers already are. Akamai has spent its life building service stations. The bet is that some of those stations can pump a new fuel without being torn down.

What The Giveback Might Be Saying

Stocks do not owe you a clean trend after good news. The giveback after the capacity announcement can mean several things at once, and it is lazy to pick only the flattering one.

It can mean fast money took the headline and left. It can mean investors want a clearer bridge from “commitment” to “recognized revenue,” because commitments are not cash until capacity is delivered and billed. It can mean worry about customer concentration: one famous lab is a wonderful reference, and also a single point of narrative risk if that lab slows spending. It can mean the rest of the business, the traditional delivery and security book, is still what the multiple is underwriting, and the market refuses to pay up until the mix actually shifts.

All of those can be true and the stock can still be the least silly way to own the theme. A buying opportunity is not a guarantee. It is a price at which the downside you can describe is smaller than the upside you do not have to invent. Under 16 times, with a signed capacity story the quote has mostly shrugged off, is closer to that description than 279 times with the good news already in the price.

A simple screen I actually use:
  Is the traffic story already in the multiple?
  Is there a contract the quote has not digested?
  Can the balance sheet fund the build without a story-stock raise?
  If the new deal slips a year, does the old business still justify the price?

Run Akamai through that screen and it does not look magical. It looks survivable. Survivable is an underrated word in a market that only wants explosive.

The Case For Waiting On The Other Two

Waiting is not the same as disliking. Cloudflare has the cleanest public evidence that agent traffic is already larger than human traffic on its network. That is a remarkable operating fact. The trouble is the embedded expectation. At a multiple near 279 times this year’s earnings, you are not buying the fact. You are buying a long string of future facts, most of which have not happened, at a price that punishes any pause.

Pullbacks in loved stocks are emotionally harder than they look on a chart. The reason you wanted in is still on television. The reason the price dipped feels temporary. That is how people pay 279 times and call it discipline because they “waited for a red day.” A red day inside a vertical move is not a margin of safety. A margin of safety is a price that still makes sense if growth merely stays good.

Fastly’s issue is different. The company is trying to move from a traditional delivery network toward an edge-cloud platform that can sit closer to AI-driven demand. The strategic sentence is fine. The shareholder experience has been a series of air pockets. Fifty times earnings is a price that assumes the repositioning works on a timetable the company can actually hit. History says the timetable slips. When it slips, the multiple does not politely wait.

So the “only one is a buy right now” line is less a medal ceremony than a relative statement. Same aisle of the market. Different asks of the buyer. One ask is patience with a full price. One ask is trust in a turnaround multiple. One ask is acceptance that a slower stock, attached to a large new commitment, may be the adult choice.

Risks Worth Writing Down Before You Click Buy

Anyone pounding a table should also list the ways the table breaks. I would rather do that in daylight.

Customer concentration cuts both ways. A flagship AI lab is the best possible advertisement and a real dependency. If that lab shifts workload back to its own clusters, or stretches deployments, the narrative premium people hope for never arrives, and you are left with the legacy multiple you already paid. That is not a disaster at under 16 times. It is a disappointment. Know which one you signed up for.

Build-out risk is next. Smaller data centers still need power, permits, hardware, and staff. The AI supply chain has been a queue, not a catalog. A commitment on paper can outrun the ability to energize racks. Delays do not erase the strategy. They do push revenue to the right, and stocks hate revenue pushed to the right.

Competition is not theoretical. The large clouds want this workload. Specialized neoclouds want this workload. The other traffic companies want this workload. Akamai’s edge is distribution and a security-and-delivery footprint that already touches a huge share of the web. That edge is real. It is not a moat you can forget to maintain. Price wars in infrastructure have a way of showing up in gross margin before they show up in press releases.

There is also the boring risk, which is my favorite because it actually happens. The core business, delivery plus security for ordinary websites and apps, could slow just as the new capacity ramps. Mix shift sounds elegant in a model. In a reported quarter it can look like “growth, but the wrong kind,” and the multiple stays pinned until the mix is obvious. If you need the stock to work in one quarter, this is the wrong setup. If you can sit through two or three muddy prints, the setup is more honest.

1. Separate the commitment from recognized revenue, and do not spend the commitment twice in your head.
2. Watch whether new capacity shows up as growth with acceptable margins, not growth at any cost.
3. Treat customer concentration as a risk factor, not as a trophy.
4. Compare any pullback in the expensive peers to their multiples, not to their stories.
5. Size the position so a delayed ramp is annoying, not existential.

How A Patient Buyer Might Actually Hold This

I am not going to pretend there is a magic allocation. People arrive with different portfolios, different tax lots, and different tolerance for a stock that can do nothing for months. What I will say is that “buy it right here” from a commentator is a view on price versus news, not a personal financial plan.

A practical way to think about it: decide what fraction of an equity book you want in internet plumbing at all. Then decide whether that fraction should be the expensive pure-play on agent traffic, the volatile turnaround, or the cheaper name with a fresh contract. Mixing a small piece of the expensive one with a larger piece of the cheaper one is a human thing to do. It is also how people smuggle a speculative bet into a sensible label. If you do it, name it. Do not call a 279-times stock a hedge.

Revisit the thesis on evidence, not on noise. Evidence would be capacity coming online, a second large customer that is not just a logo, security and delivery revenue that does not roll over while you wait, and margins that do not collapse to win the AI work. Noise would be a rival’s keynote, a one-day move in a basket of infrastructure names, or a social-media thread that discovers edge computing for the third time this year.

Dividends and buybacks, if they continue, are the quiet part of the Akamai story that momentum accounts ignore. You do not buy this for the yield. You also should not ignore that the company has a history of returning cash, which is a different posture from a story stock that needs the market to fund every new chapter. Cash return does not make the AI plan work. It does lower the cost of being early.

What “Getting The Partnership For Free” Really Means

The phrase is punchy, and punchy phrases deserve a translation. It does not mean the partnership has zero value in the model and you are being handed a call option. It means the stock, after giving back the initial jump, is not priced as if that commitment is the center of the equity story. The multiple still looks like the multiple of the company people thought they knew.

If the commitment converts into capacity and then into revenue at reasonable margins, the multiple can rise and the earnings base can rise. That is the double that value-oriented growth investors actually want, and rarely get both of. If the commitment stalls, you are roughly where you started: a sub-16-times infrastructure name with a slower year than its peers. The asymmetry is the appeal. Not a lottery ticket. A skewed range of outcomes.

I have found that investors talk themselves out of skewed ranges because the upside is not cinematic enough. A stock that might re-rate from the mid-teens to the low twenties on earnings, while earnings grind higher, will not trend on social media. It can still beat a stock that needs perfection and delivers very good. Very good, at the wrong price, loses to good at a fair price more often than the highlights suggest.

Agent Traffic Is A Volume Story And A Mix Story

There is a trap in treating all new requests as equal. Some agent traffic is valuable: authenticated, tied to a paying workflow, willing to sit behind a commercial contract. Some of it is noise: indiscriminate fetching, training scrapes, retries that should have been cached. The companies in front of the web do not get paid the same rate for both, and they do not want both in the same proportion.

That is why the security and filtering layer matters as much as the raw delivery layer. If agents flood a site, the owner of that site will pay to sort the flood, not merely to carry it. Sorting is a better business than hauling. Akamai, Cloudflare, and Fastly all sell some version of sorting. The one that converts agent chaos into paid policy, rather than into unpaid bandwidth, keeps the economics.

Watch the language on earnings calls for this. If management talks only about volume, be suspicious. If they talk about which workloads they will refuse, price differently, or push to the edge, they are managing a network, not renting a pipe. Networks that cannot say no eventually discount themselves.

Useful question for any traffic stock: is the new volume priced, filtered, and close to the user, or is it just more bits?

A Longer Look At The Seven-Year Shape

Seven years is an awkward horizon for public markets, which often cannot see past the next two prints. It is a natural horizon for capacity. You do not stand up distributed compute on a quarterly cadence and call it strategy. You sign power, you place hardware, you hire operators, you migrate workloads, and you hope the customer’s own product cycle still wants the capacity when the racks are warm.

Spread across seven years, $11.6 billion is a program, not a spike. Programs get revised. They also anchor internal capital plans in a way a one-off purchase order does not. For a company trying to convince its own engineers that the edge-inference bet is real, an external commitment of that size is internal permission. That sounds soft. It is not. Companies abandon half-built strategies when the external world will not co-sign them. This one got a co-sign.

The stock market’s job is to doubt the co-sign until revenue shows up. Your job, if you are the buyer the commentator described, is to decide whether you want to own the doubt at a mid-teens multiple or own the belief at a triple-digit one. I know which of those lets me sleep. Sleep is not a research method. It is a useful tie-breaker when both stories are plausible and only one price is forgiving.

Where The S&P Comparison Helps And Where It Lies

All three names have beaten the index this year, and that fact gets used as proof that the theme is working. It is proof that the theme has been bought. Beating the index is not the same as being attractively priced for the next stretch. A stock up 159 percent can be a wonderful company and a poor entry on the same afternoon.

Akamai’s 22 percent gain looks tame next to its peers and still beats a roughly 12 percent index move. Tame is the feature. You are not being asked to chase a vertical line. You are being asked to notice that a large contract landed on a line that never went vertical. In a year when plenty of AI-adjacent names have gone vertical and then sideways, tame-plus-contract is a combination I would rather underwrite than vertical-minus-margin-of-safety.

Index-relative charts also hide business quality. The index is a blend. These three are specific. If agent traffic fades, the index will barely notice and these stocks will. Size the exposure like a specific bet, not like a substitute for the market.

What Would Change My Mind

A view without a kill-switch is a hobby. Here is mine for this setup.

I would get more interested in the expensive peer if the multiple compressed for a boring reason, growth still intact, rather than for a broken one. A pullback that leaves the operating story alone is exactly what the commentator said he is waiting for. I would get less interested in Akamai if the capacity program slipped without a clear operational cause, or if margins on the new work looked like a land-grab that never ends. I would get more interested if a second independent customer signed something you could measure, because one lighthouse account is a start and two is a market.

I would also watch the language around smaller data centers. If the plan stays specific, locations and workloads and payback, that is execution. If it drifts into a cloud-vocabulary fog, that is marketing catching up to a stock move that has not happened yet. Words are cheap. Energized sites are not.

The Human Habit This Trade Is Really About

There is a habit, and I have it too, of equating the stock that moved the most with the business that is winning. Movement is a record of who showed up to buy, on what story, with what time horizon. It is not a record of who owns the scarce asset. Sometimes they overlap. This year, in this little trio, they have overlapped unevenly.

Fastly moved the most. Cloudflare has the loudest traffic anecdote and the richest multiple. Akamai signed the capacity commitment that actually sounds like the next phase of AI, inference in more places, and then watched the stock give the celebration back. If you are building a watchlist rather than a victory lap, that last sequence is the one to sit with.

You do not have to love the commentator to use the frame. The frame is old. Buy the version of the theme the market has not finished celebrating. Ignore the version that needs you to believe in a straight line. Recheck both when the facts change. Most people invert it, because celebration is easier to see than absence.

Putting The Buy Case In One Place

Strip the television cadence away and the argument is short enough to test.

AI agents are already a visible source of web traffic, large enough that at least one major network has said agent traffic passed human traffic. The firms that manage that traffic have all beaten the broad market this year, but the market has paid them very different prices. A new, seven-year, $11.6 billion capacity commitment landed on the name with the lowest multiple and the mildest year-to-date gain. The initial price jump faded. The commitment did not. At less than 16 times this year’s expected earnings, the stock still looks like the old business, while the contract belongs to the new one.

That is the whole pound-the-table case. It does not require you to think Akamai will out-innovate every cloud. It requires you to think the market is slow to refile a company it has known for a long time, and that a slow refiling at a sane multiple is a better entry than a fast refiling at a wild one.

I think that is right, with the caveats above taped to the screen. Right is not the same as urgent. Urgent is how people overpay. The quote already cooled off. Cooling off is the condition the call was waiting for.

Questions Worth Asking Before The Next Print

If you own it, or you are about to, the next earnings conversation should be narrower than “how do you feel about AI?” Feelings are fully priced in this sector. Specifics are not.

• How much of the capacity commitment is scheduled for the next twelve months, versus the back half of the seven years?
• What does incremental margin look like on edge inference relative to classic delivery?
• Are agent-driven requests being filtered and priced, or simply carried?
• Is power, not demand, the binding constraint on the smaller halls?
• Does management still sound like an operator, or has the script drifted into theme language?

You will not get crisp answers to all five. You will get a tone. Tone plus one hard number is usually enough to tell whether the program is moving or being narrated.

A Note On Temperament

Infrastructure stocks reward a temperament that momentum markets punish. They gap on contracts, then bore you. They miss a quarter on timing and get treated like the strategy died. They work, when they work, because the map and the contracts compound while the commentary rotates to whatever doubled last month.

If that boredom feels like a problem, this is not your stock, even at a pretty multiple. If boredom feels like the price of not overpaying, Akamai is built for you in a way the other two currently are not. I would rather be slightly early and slightly bored than precisely late and fully excited. Excitement is expensive this year. Boredom, attached to an eleven-figure commitment the quote has mostly ignored, is on sale.

None of this is a promise that the next buyer will agree with you on your timetable. It is a description of a gap. Gaps close. Sometimes they close by the story failing. Sometimes they close by the price catching up. The work is deciding which failure you can afford. At this multiple, with this contract, I can afford the first more easily than I can afford the prices on the peers.

So the line that stuck with me was the plain one. You can buy that stock right here. Not because the internet is about to be rewritten in a keynote. Because the rewriting is already showing up in traffic logs, because someone just agreed to pay for capacity over seven years, and because the share price, after a brief cheer, went back to acting as if little had changed. Little had changed in the quote. Quite a lot had changed in the order book. That is the discrepancy. Discrepancies are where long-only investors still get paid, when they are willing to look slightly out of step for a while.