An old friend wrote on LinkedIn about there being no way that AI costs could potentially 10x as I shared a couple days ago.
“Amazon was not profitable the first 8-9 years in business. It’s not all about profit now, but establishing brand and a future customer base that will stay with them as they grow. The prices will not go up 10x. They already have price tiers for people who need the extra power.”
Here’s why I think he’s wrong. The Amazon comparison sounds great until you look under the hood.
1/1 Amazon scaled efficiency within existing infrastructure already in place. They centralized hubs and spokes for distribution and they leveraged the internet as a single storefront, all while niche competitors, small mom and pops and everyone in between were slow to the draw and slow to adapt.
The direct to consumer sellers had their own costs that they couldn’t get away from while people used their businesses as showrooms only to later go to Amazon to purchase their goods when they learned that their size 9 shoe was the size they needed. Amazon took links out of the supply chain, and every link they removed was someone else’s profit gap. Amazon’s economics improved with every additional customer because every customer made the machine cheaper to run per transaction. And let’s not forget, Amazon also had all the benefits of existing infrastructure and they also used and abused USPS with us, the taxpayers, subsidizing USPS losses. They rented space until they could buy and build their own. Amazon’s only real scarcity was time.
AI is actually the inversion of the Amazon model and you’re missing it completely.
Every additional user makes the AI machine more expensive to run. More inference means more compute -> more GPU -> more power -> more cooling (more power) -> more data centers (more power and resources to run them). The marginal cost doesn’t trend down with AI. It trends up. Amazon removed friction from a system while AI adds load.
Now layer on external constraints Amazon never faced.
1) Physical scarcity. TSMC is a massive bottleneck for advanced chip fabrication (and also a national security risk with everyone relying on them and their currently being on China’s doorstep and one boom away from, “Oh shit!”. It will be at least 3 to 10 years before we have control of manufacturing capacity and it catches up with current demand. And demand is growing! We can’t spin up GPU production the way Amazon could lease warehouse space. And beyond manufacturing, there are mineral realities to contend with. We know the resources exist on our own land, but bureaucracy and environmental concerns keep us from efficiently extracting them while China holds critical supply chains hostage.
2) Energy. The grid is already taxed and aging. We don’t even manufacture the transformers and switching equipment we’d need if any of our major stations had an issue or outage. And we know data centers are enormous power draws hitting a grid with little room to give. Our grid is already so overtaxed so power generation needs to happen almost on premise. That’s a lot of solar and a lot of batteries until we get nuclear plants or small modular reactors dialed in. So now we’re trying to fire up old decommissioned reactors and build datacenters on their doorsteps. Great, but that also comes with its own host of regulatory, safety, and timeline issues. Amazon never had to navigate and negotiate the power constraints that AI is creating – even with AWS – but AI is creating some nightmares right now even for them. But it didn’t affect their shipping boxes.
Now you might bring up quantum computing as a savior. But quantum doesn’t do what AI needs, quantum doesn’t run inference. Quantum excels at narrow problem classes like optimization and molecular simulation. AI runs on massive-scale matrix multiplication across neural networks. That’s classical compute. And even if quantum eventually contributes, it comes with its own version of every constraint I just described. It’s not cheap to cool anything to near absolute zero, plus all the rare materials, yada yada, yada. So quantum doesn’t solve the problem. It’s just a different problem..
So that brings me to the business model in general.
Your point about price tiers actually supports what I’m saying. Those tiers exist to hook people in but even flat-rate pricing is unsustainable. It’s why a $200/month Claude Max subscription running heavy workloads is still dramatically cheaper than direct API access. The API costs are more realistic, but they’re still heavily subsidized with investor cash. A 5-8x increase for comparable capability is entirely defensible based on current burn rates, and even that doesn’t service the debt these companies continue to accumulate.
Then there’s zero brand loyalty! None. OpenAI is trying to build it with ChatGPT and cheap subscritpions, but as soon as people discover how much better Claude is they jump, until next week, and the game of hopscotch continues back and forth. That means these companies can’t slowly raise prices without hemorrhaging users, which will make the correction sharper when it finally hits.
The companies subsidizing your AI usage today are burning capital at historic rates with no clear path to profitability, at least not at current pricing. That’s not a business model. It’s a ticking time bomb.
And this is why I make the argument for taking control of your own AI infrastructure. Plus the privacy and security factor. Invest in local GPU horsepower. Learn to run and quantize open-source models for your specific use cases. Think about energy independence. Stop assuming current pricing is permanent.
There is one path out: the models themselves have to get radically more efficient. Smaller, specialized models instead of massive general-purpose ones. Mixture-of-experts architectures that only activate the parameters they need. Distillation. Better inference optimization. We’re seeing early movement here, but the gap between where efficiency is and where it needs to be to support current pricing is enormous. And that gap is growing faster than the efficiency gains.
We need to make AI useful and profitable, now, while the window is open and build enough margin to cover the real costs when they arrive.
If you or your company need help bringing services on-premise, reach out. I may be able to help or point you in the right direction.
