Dennis Yu on Anthropic's Super Bowl Stunt and the AI Pricing War Small Business Owners Should Be Mining

Featuring Dennis Yu, founder of BlitzMetrics and a longtime search and AI engineer.

Anthropic's Super Bowl ad against ChatGPT looked like a values campaign. Dennis Yu told me it was an architectural confession dressed up as a moral one. While the foundation models burn cash to lock in users, the people who quietly win this round are small business owners who buy the right tier and turn AI workers into business assets.

I sat down with Dennis after dinner at a Chinese buffet in Alpharetta, Georgia. We had been at Capital City Roofing earlier that day with Brad Strawbridge, watching a roofer use Antigravity to spin up a new website in three days. So when we started talking about Anthropic's Black Mirror–inspired Super Bowl ad, the conversation kept circling back to the same point: the model wars are loud, but the leverage for small business owners is somewhere else entirely. Dennis has been a search engineer for more than two decades and runs the Dennis Yu YouTube channel where he teaches the Marketing Mechanic series. I wanted his read on the news cycle and what it actually means for someone running a real business.

Read Anthropic's Super Bowl ad as a confession, not a values statement

Anthropic ran a Super Bowl ad that lifted a premise from Black Mirror, a hospitalized woman who has to pay extra for the ad-free version of her own brain function, and pointed it at OpenAI's plan to put sponsored results inside ChatGPT. The marketing students in Dennis's senior seminar at his university could not stop talking about it.

"I think it's deliciously ironic because Anthropic is one of the most misanthropic names on the planet."

Dennis's read is that Anthropic is the company that paid the $1.5 billion settlement for ingesting copyrighted material, while also suing model-condenser startups for synthesizing Anthropic's rules. The Super Bowl ad worked because most of the audience does not know that. The reaction in Dennis's classroom mirrored exactly what the campaign was engineered to produce.

The lesson for me as a marketer is not that the ad was wrong. It is that a tightly engineered campaign can move a room full of marketing seniors to the exact conclusion you scripted, even when the brand making the argument has the weaker hand. Anthropic accomplished what they wanted to accomplish. Sam Altman replying with a five-paragraph essay was the tell.

The DoD contract reveals an architectural truth Sam Altman won't say out loud

The story most people missed is that Anthropic holds a roughly $200 million per year contract with the Department of Defense, and Anthropic models helped surface the intelligence that led to the capture of Nicolás Maduro in Venezuela. Dario Amodei's public answer for why Anthropic can take Defense money but not enable autonomous targeting or surveillance of American citizens is that the company "draws the line" on values.

Dennis pushed back on that framing hard:

"Sam Alman can do that deal because he doesn't specify the rules ingrained down here. The reason Anthropic can't do this deal is because they can't change what's already — they'd have to retrain, rebuild their entire platform to do that."

OpenAI's values sit in a rules layer above the model. Anthropic's values are baked into the training itself. That is why Altman can sign deeper Defense contracts without retraining a thing, and Dario cannot. It is an architectural constraint sold to the public as an ethical posture.

The surveillance-versus-targeting distinction collapses the same way once you look at how the data moves. To decide whether a target is an American citizen, the model has to ingest the person first.

"Without the data, without labeling, you have to have ingested them in order to label them, right? Like there's no — you don't have to be a data engineer to know that that's the case."

If you have the data infrastructure to identify Maduro from foreign signals, you have the data infrastructure that could identify a Texas plumber. The line between "we surveil foreign nationals" and "we surveil Americans" is a UI choice, not a capability boundary.

Stop chasing this week's model release

Every week one of the foundation models leapfrogs the others. ChatGPT releases 5.5 and the internet declares it the new winner. The next week Gemini ships an update and the conversation flips. Dennis is openly bored by this cycle.

"These guys are all kind of circulating against each other. They're all moving in the same direction. We know what that path is. We know where that's going for the next 36 months."

His point is technical. An LLM is a word probability calculator with three levers: the input (context window, multimodal inputs, connectors), the processing layer (the model itself), and the output (more agents, more tools it can drive). All the model labs are pulling the same three levers because there are no other levers. Whichever one ships a feature first sets the agenda for a week, then the others copy it.

If you understand that, you can stop reading the launch posts. You can also stop pushing your team to migrate stacks every quarter based on a benchmark.

Memory is OpenAI's switching cost, not a feature for you

Here is the most useful framing Dennis gave me on the consumer side of the war. OpenAI has roughly a billion users, somewhere near seventy-eight percent of the world's LLM users. That base naturally pulls them toward an ad-supported model, and toward making the cost per inference as cheap as possible. The 5.0 launch, in Dennis's read, was less about a smarter model and more about wringing more economics out of every token.

To keep those users from drifting to a better model, OpenAI needs lock-in. The cleanest lever they have is memory:

"They have over prioritized memory to increase the switching cost."

Dennis frames it like a relationship. Once ChatGPT knows your kids' names, your business, your tone of voice, and your last six months of decisions, you are not going to dump it for the next hot model even if the next hot model is sharper. It is the loyal girlfriend problem.

Anthropic and Gemini are not playing that game because they are not chasing your consumer mindshare. Anthropic is chasing the business user the way Microsoft once chased SQL Server seats while Apple chased the consumer. The product surface tells you who the customer is. Anthropic optimizes for longer task completion, longer state continuance, and the ability to see a business job through to a result. OpenAI optimizes for "remembers you."

If you are a business owner, the question is not which model has the cuter memory. The question is which one your team can run a real workflow inside.

The real bottleneck is electricity, and Elon is the only one acting like it

Dennis kept pulling the conversation up the stack. Above the SaaS layer sits the agent layer. Above that, the models. Above that, the chips. Above that, the data centers. And above that, the actual electrical grid.

"The limiting factor right now is physical electrical."

Jensen Huang, Sam Altman, the Google chairman, Bezos, Dennis kept noting that every major operator says the same thing in public. China is operating at roughly three terawatts and still building nuclear plants. The US is at about one terawatt and the build-out has been slow. Meta is setting up power plants. Elon stood up Colossus 1 outside Memphis and is building Colossus 2. Google has patents for floating data-center containers parked offshore.

The first-principles answer Elon is betting on is to move the data center to space. No weather disruption. Free cooling on the back side of the panel. Solar power that never sleeps. Starlink as the downlink. SpaceX bought xAI partly to put that bet on the same balance sheet.

You can be skeptical of the timeline. I am. But the strategic point holds that the AI race is going to be won by whoever solves the electricity problem first, and the consumer-facing model wars are downstream of that fight. When Nvidia announced a $10 billion AI investment instead of the rumored $100 billion, both Nvidia and the partner stock took a hit. Wall Street is starting to price the bubble risk because the capital commitments only pencil if the power can actually come online.

This is a once-in-a-decade pricing war… take the free money!

Here is the part of the conversation that matters most for the reader of this article. While the giants fight, they are subsidizing the tools.

"I burned 150 million tokens in the last week. Okay. I paid $200 per month. So, I paid $60 to burn 150 million tokens. And for those who don't know what that means, that's like $6,000 of processing."

Claude is selling roughly $6,000 of compute for $60 of plan time because they are losing money on purpose to take market share, the way Uber did against Lyft for a decade. Perplexity is doing the same. ChatGPT subsidizes the free tier through ad revenue and a billion users.

Dennis's frame for who should pay what is the cleanest I have heard:

• If you have no money or you are a student, use the free tier of ChatGPT and accept the ads.
• If you can spend twenty dollars a month, ChatGPT Pro or Claude Pro is fine.
• If you are a serious business owner who can spend at least an hour a day inside the tool, buy Claude 20x Max at $200 a month.

"It's like, imagine having six people working full-time for you. It's not a lot of money when you actually use it."

He added a caveat I want to repeat. Free value is not free if it sits unused.

"Just cuz something is 95% off doesn't mean it's a good deal unless you use it."

If you buy the $200 plan and use it like the $20 plan, you are losing money. Match the tier to your actual workload.

Build capabilities, not skills

The last big idea Dennis dropped on me is one I am still chewing on. I had told him that the real win for the consumer in this pricing war is that you get to learn a transferable skill across all the models. He cut me off.

"A skill I think is typically tied to a person... A capability is something that's the property of an asset."

His point is a skill lives in a head. If you go on vacation, the skill leaves with you. A capability lives in the business. It is process, system, and orchestrated AI workers, Dennis literally names his agents Jennifer and Ethan, that operate whether the owner is on the operating table or hovering ten feet above it.

"If you own a business, forget the vanity and like you're working in the business... In order for you to have ongoing value to scale your business, you have to be able to extricate yourself from the business."

This is why Dennis is unimpressed by founders who can run a flashy demo but cannot leave their business for a month. The right question is not "which prompt did you write?", it is "which bottleneck did you remove from the org?" Feed your QuickBooks, HubSpot, and Slack into the model and ask it to find your bottleneck. Then build the capability that fixes it.

Brad Strawbridge built a roofing website, a licensing-model website, and a nonprofit website in three days using anti-gravity. He is not an AI engineer. He told us how he did it: he asked the model what to do, and when it said "log into your GitHub," he asked, "What is GitHub?" The capability he built was the asking-the-model habit. Read our conversation with Brad on AI in the trades if you want the full picture.

What small business owners should take from Dennis's playbook

If you take three things from this episode, take these.

• Pick whichever model you'll actually use, then stop comparing. The consumer-tier LLMs are Toyotas and Hondas — windshield wipers in a slightly different place, same car. Commit to one for ninety days, get fluent inside it, and ignore the launch posts. The cost of switching is one hour. The cost of perpetual A/B testing is your year.
• Match the plan to the workload, not to the FOMO. The $200 Claude Max plan is only a steal if you are spending an hour a day inside it. If you are not, the $20 plan or the free tier is right for you. Free compute that you do not use is the most expensive line item on your P&L.
• Build capabilities, not skills. Name your AI workers. Document the process they run. Make the business operate without you for a week. If you cannot, you do not have a business — you have a job that pays inconsistently. The pricing war gives you the cheapest opportunity in a decade to fix that.

Watch the full episode

Watch the full conversation with Dennis Yu on YouTube: the full episode is here. If you want help wiring AI workers into your operations the way we do for service businesses, book a Nail demo .

Article by George Paladichuk, founder of Nail AI. Featuring Dennis Yu, founder of BlitzMetrics and host of the Marketing Mechanic series on the Dennis Yu YouTube channel.