Disclaimer: I am not a trained analyst, I am biased and hope I'm right, and I heavily use AI to assist in research as well as to red-team theses. NFA/DYOR. Position: 1100 shares @ 25.59.
Let's cut right to the chase
Comcast has been battered. At ~$22/share, it hasn't been this low since 2013. That is not unwarranted. Comcast is a broadband provider, and customers are being poached left and right by fixed wireless (FWA) with the newly added anxiety of having to compete with Starlink and AST SpaceMobile. Its core cash cow is bleeding subscribers. These things are all true, but it's too early to declare time-of-death on broadband — the crux of this bull case rests on the coming macro shift of the AI industry from training to inference. As AI becomes more integrated in our daily lives, as open-weight models close the frontier gap (and Fable 5 has shown why reliance on a central cloud is risky), and as appetite for tokens and power grow, Comcast will be ideally positioned; it already possesses a distributed network of mini data centers touching 65 million homes as well as demonstrated proofs-of-concept, in partnership with Nvidia and others, to monetize it. Stay with me as I show why the bottom is in for Comcast and, in addition to a great value play, make a case for growth as well.
The Floor
I'll get the boring stuff out of the way because it's important to anchor the rest of the conversation, so bear with me (haha) for a moment before we talk about the juicy stuff. The bear case is straightforward: Comcast's broadband infrastructure, a mix of coaxial and fiber optic cables to deliver internet to homes, is under siege from fixed wireless and satellite. Expensive to maintain compared to cell towers, price-conscious customers are cutting the cord. The last four quarters saw a total of over half a million subscribers ditching Comcast as their ISP. If this continues, there will be nobody left to pay the bills. Plus, with AI euphoria in full swing, value names are boring as investors chase returns elsewhere.
Those losses, while concerning, appear to be decelerating, down from 226k in Q2 2025 to 65k last print. The mirror image of that is Verizon's FWA gains, down 27% YoY, coupled with its increased emphasis on fiber. Some analysts even see broadband actually returning to growth in 2026. The Q2 earnings report will confirm the trend or invalidate it. Still, worst case, against a total user base of over 31M subscribers, half a million per year is less than 2%.
There is also evidence that the piranhas are getting full. According to New Street, there is only so much FWA capacity that the current infrastructure can serve. There is only room for about 17M more households, full-stop, on fixed wireless. So pretending that it can completely displace broadband is ridiculous.
The fundamentals at this price are flirting with absurd territory. The shares are trading at a price to earnings multiple of ~5x. The dividend is $1.32/share, which corresponds to a 5.9% annual yield against $15.6B of free cash flow, a 32% payout ratio, which means even such a massive yield is extremely safe. The company's authorized share repurchase allotment is $15B, of which it is deploying roughly $1.5B/quarter. This would reduce the float by ~18% over 2.5 years, massively juicing earnings per share and returning value to shareholders above and beyond the already attractive dividend. Taken together, that means a 13% annual return to shareholders irrespective of any of the bull catalysts we're about to talk about. As for debt, 2.3x net leverage is comfortable.
Just to be thorough, if you place any stock in analyst price targets, the upside implied by the average is 47%, with even the hungriest bear suggesting a floor at $21 (-5%).
The Bear Defense
As it happens, Comcast is wielding a potent canister of bear spray. The first arrow in the quiver (to mix metaphors) is DOCSIS 4.0, custom silicon developed in conjunction with Broadcom that brings coax up to par with fiber. Instead of the enormous expense of retrenching cables, DOCSIS 4.0 delivers comparable performance at a fraction of the cost by allowing the lines to run full-duplex, greatly extending coax's lease on life.
As for stemming the broadband hemorrhage, Comcast is aggressively pushing wireless/home internet bundles and offering same-day Wi-Fi connection, eliminating a major cause of retail friction.
It is also more than just broadband. Comcast owns Sky (a European news conglomerate), the Peacock streaming platform (Netflix's business model), and a plethora of NBCUniversal theme parks and concomitant IP (Disney's). Why is a company, playing the same game as News Corp (33x), Netflix (25x), and Disney (16x), priced like none of them? The answer is probably that these revenue streams make up approximately 27% of the total, the majority being "dying" broadband, but it's still a splitscreen worth considering.
Alright, that's the end of the boring part. Hopefully we've established that this is at worst a value name priced for annihilation by a combination of bad narrative and more seductive gambles in the growth sector. What remains is a free call option bolted on top, and you are paid handsomely waiting for it to pay off.
Offense
The single most important macro-level wave to consider is the shift from AI training to inference at the edge. This is the moneymaker. By 2030, inference will be ~70% of demand, up from ~55% today. My bet is that given the constant need for power, tokens, and freedom from executive order, a healthy portion of that will be at the edge.
You don't get it there over wireless. Wireless is like sending smoke signals over the horizon, while broadband injects data via 14-gauge bore straight to the jugular. The physics can't compare. Latency and throughput will count increasingly more as customers' appetite for tokens only grows.
At the heart of Comcast's AI strategy is a three-legged stool of AI-enabled advertising, an AI "concierge" for SMBs, and low-latency inference for gaming. I've said that Comcast possesses a "distributed data center" that touches 65M homes, and that's the moat. While hyperscalers are arguing over permits and waiting in grid queues, Comcast's existing footprint consists of ~200 regional data centers, 1,300 vCMTS nodes, and 50k+ edge servers throughout the majority of the US. Comcast hasn't said exactly how they're serving all that inference, but based on announcements from Charter (which runs a very similar playbook), we think it likely that its nodes are already being packed with Nvidia's RTX PRO 6000 Blackwell Server Edition cards or equivalent. Here are some proven ways this is already paying off. Proven, but still early. First, household-level personalized ads driven by AI in partnership with Decart (real-time AI video). This one is in the trial phase. Second, an AI concierge for small-to-medium-sized businesses. This one is live on the grid today. Third, the most proven of them all, is low-latency inference for gaming. Currently deployed Nvidia Reflex and Rivermax packet pacing for GeForce NOW delivers faster performance than a local PS5. If that weren't real enough, Comcast's hiring telemetry corroborates: there are just under 100 job listings currently open for machine- or deep-learning engineers, including one that specifically asks for "ML inference optimization."
The last thing I'd like to touch upon is the power physics. Comcast's distributed network already does a good job diffusing power requirements throughout the grid and bringing high-bandwidth, low-latency data right to the home. There is another layer that takes this to the next level, and it involves passing the power burden onto the consumer entirely. This is real and early. Comcast Business and Dell operate a "managed far edge" service on Dell PowerEdge servers and the NativeEdge platform to run inference locally at the customer's location. Imagine a hospital that doesn't want PII leaving the premises, or a bank that would prefer to keep its financial data in-house — not to mention all the power users that still have PTSD from Fable 5 being ripped away from them mid-prompt. The best part — Dell NativeEdge is hardware-agnostic, meaning when custom inference silicon displaces Nvidia for a particular use case, there is no reason not to switch. The data already support inference taking an increasingly large slice of the pie going forward.
Lessons from the Fable 5 Debacle and the Real Bull Case for Comcast
A quick recap for the uninitiated. Throughout 2026, Anthropic was hyping its most powerful model, Mythos, as a veritable cyberweapon and giving select enterprise customers early access so that they can shore up their attack surfaces. In the early days of June, word got out that the consumer would soon get access to a Mythos-class model. On June 9, it happened. Fable 5, Mythos but with cybersecurity and biology guardrails, hit retail accounts everywhere; and for about 72 hours, it was glorious. Then Anthropic received a letter from Howard Lutnick, Trump's Commerce head, informing them that, due to an alleged successful jailbreak, Mythos and Fable had been classified under export controls for non-US citizens and giving them 90 minutes to comply. Having no way to differentiate at the user level, Anthropic had no choice but to pull access to Mythos/Fable entirely.
So what does this have to do with Comcast? This is where we get speculative. The answer is sovereignty. Customers don't want to live in a world where their biggest engine of productivity can be yanked from their grasp by the government and/or a private company without warning. Customers likewise want to be able to work on cybersecurity/biology-related topics without getting an earful from their AI overlords. They might also need to work with sensitive data, such as medical/financial information or API keys, without letting it leave the local network. All of these reasons will force companies to consider whether they want to trade a little bit of capability for sovereignty and privacy, and in my opinion, it's hard to imagine that there isn't a significant chunk of end-user demographic that will choose to make that tradeoff. Schools, hospitals, banks, police stations, burger joints, and vibe coders everywhere will soon have two main options: A) keep paying API rates for centralized models that hoover up all their activity and could be turned off at any time, or B) self-host and avoid all that (before you yell at me, I know that's a false dichotomy; they could simply switch providers or use tools like OpenRouter — but that's just trading one single point of failure for another).
I will caveat by saying that this prediction is based on open-weight models maintaining near-parity with closed, which is admittedly uncertain. Open weight could stall here, or closed source could compound too fast for open to keep up. However, that hasn't been the case so far. GLM-5.2 is currently second only to Fable on Arena AI's web development benchmark, with some predicting parity with today's Mythos by early 2027.
But there's a problem. The most capable open-weight models require amounts of compute that you don't find in a gaming desktop or even most commercial server rooms. You need dedicated clusters approaching data-center-level orchestration to run multiple instances. Enter Comcast.
Why Comcast?
The short answer is it's not just Comcast, it's broadband. Written off by the market, I believe broadband will see a miraculous comeback as edge inference needs increase because A) FWA/satellite cannot compete on latency/throughput, and B) data sovereignty will become increasingly important to the user. Most people already lease their router for simplicity. Why not their inference clusters too?
Here's an example. Consider the residential user that spends most of his time maintaining his GitHub or gaming. In an owned-inference model, the leased node could serve 90% of his daily inference needs. However, since he is not an enterprise customer with a dedicated server room, the largest, cutting-edge models would likely require an API to a central provider — unless he goes through his ISP. In that case, the entrenched network shines. The regional nodes serving medium-sized geographical areas could, in conjunction with all of the other services already described, be responsible for handling the heaviest calls. An integrated network would have no trouble elevating to the smartest models when necessary while delegating smaller tasks to the lowest level available, all without noticeable lag.
I'll be real with you. It doesn't take a huge technological moat to set up a cluster capable of running open-weight models and plugging in your MCP. Dell, Hewlett-Packard, and others are more than capable of doing so, and even the customers themselves, if they have a competent IT team, could probably figure it out. AT&T or Verizon could just as easily roll a truck to your house and bolt a couple of GPUs to your wall. There are no 80% margins here for Comcast to perform this service. What it does require is a large, coordinated, integrated network, the kind of thing that broadband already is and the only thing that hyperscalers cannot replicate. My bet is that lightning-fast, sovereign inference will mandate distributed and coordinated power as well as existing customer relationships. All of that spells "broadband." There are many broadband providers to choose from, but Comcast is the one I own. This analysis is probably portable enough between them. In my view, it's the cleanest expression of that thesis today.
Numbers
We have described the thesis qualitatively, now let's try to quantify it. The numbers vary wildly depending on whether broadband, the transport layer for the coming inference takeover, captures any value. At worst, we could see a return to broadband growth. That alone would constitute a reversal of current trends for the entire sector. At best, broadband could go from "dumb pipe" to "smart, integrated router" for all your inference needs, handling seamless orchestration for the customer while the hardware providers fight over market share. Imagine for a minute that you are a customer interested in self-hosting a fleet of 1.5T open-weight models for your home or business. Having the ISP you already pay handle everything would be simpler, safer, and cheaper. One phone call or a few taps in the app and you can upgrade your account to AIaaS. The same day, Comcast arrives with a plastic, weather-proof box that sits next to your gas meter, comes equipped with the best models (and upgrades them as they are released), and is ready to operate agentically within your network in a manner that is not vulnerable to a sudden executive order, all for a reasonable monthly payment.
That payment would obviously differ significantly if you are a single-family household or a multi-jurisdictional business. For reference, Comcast's broadband revenue is approximately $26B (72%) from residential and $10.2B (28%) business. Enterprise will face greater pressure to convert due to competition, continuity, and data privacy considerations. I think it is reasonable that about half of Comcast's residential accounts and approaching 90% of its business accounts seek to upgrade their accounts to AIaaS, but I'll run the numbers in a low-adoption scenario to establish the base case. If only ~10-25% of residential customers convert for an additional $10-15/month and 50% of enterprise for +15-25% of their current bill, that would translate to ~$1.1-2.7B of incremental revenue, to say nothing of pure subscription count reaccelerating. While it may not sound like much in absolute terms, the growth potential is what justifies the multiple.
The astute reader might be wondering about equipment depreciation and/or availability of GPUs/ASICs for inference. Happily for Comcast, inference is much more forgiving on both fronts than training. It is the training silicon, such as the H100/H200 or the GB200/B200, that are currently being fought over as market pressures necessitate the cutting edge at all times. When it comes to inference cards, such as the RTX PRO 6000 Blackwell namedropped by Charter, they are much more plentiful and long-lived.
Speaking of multiples, I believe that a reversal of the broadband narrative alone warrants a rerate to around 10x — approximately double from here. The bull case is how the rest of the thesis plays out. The AI enablement programs are in the sweet spot, between having been proven in field trials and rolled out in limited capacity and showing up in the earnings transcripts as real revenue. If that story materializes, expect a multiple between 10-15x. If Comcast turns out to be the winner among broadband/edge inference providers, it could justify 15x+.
TL;DR
Broadband providers in general, and Comcast in particular, have been unfairly punished by the market due to fears of obsolescence at the hands of FWA and satellites. This is overblown, and today's valuations already represent an attractive opportunity to say nothing of growth potential.
The growth potential is very real, and it rides on the back of a market-wide transition from AI training to inference, and among that inference, edge compute. Broadband providers are uniquely positioned to capitalize on this trend due to their existing distributed networks and customer relationships. Just like people currently rent their routers from their ISP, they will rent local inference clusters to serve their in-house needs, elevating to regional nodes where appropriate. The providers of the silicon will be in competition with one another for selling the best hardware, but broadband is the entrenched last-mile that will benefit from the competition to lease hardware to the user.
Personally, I'm scooping up shares at depressed multiples because I expect a rerate when this becomes clear to the market. Don't do the same because I said so, but let me know if you agree with me.