I’m Buying Every Share I Can (Investors Aren’t Ready)

Table of Contents

1. Introduction
2. AMD's Bet on Hyper‑Low‑Cost Chips
3. Cerebris's Wafer‑Scale Approach
4. Nvidia's Asset‑Class Vision

Key Takeaways

• AI infrastructure spending is accelerating, not slowing, despite negative free cash flows at hyperscalers.
• AMD, Cerebris, and Nvidia each pursue radically different hardware strategies for the next phase of AI.
• Data center resources (land, power, cooling) are zero‑sum, ensuring continued investment shifts rather than a slowdown.

The single biggest argument against AI stocks just died. Wall Street analysts have been saying the same thing for years now. AI is a bubble and the entire tech sector will crash as soon as spending slows down even a little. But three major developments just put an end to that argument and kicked off what could be the next phase of the AI era. My name is Alex and I spent eight years as an electrical engineer and AI researcher at MIT. And I've never seen chip companies move this fast. Let me show you what's going on and how I'm investing in it. Your time is valuable, so let's get right into it. The biggest case against AI has always been that spending will eventually slow down. And when it does, every stock that lives and dies by that spending will go down with it.

It's only a matter of time. Hyperscalers like Google, Amazon, Microsoft, and Meta platforms report their CapEx budgets every quarter. And every quarter, it's been growing like crazy. In fact, their AI spending has been compounding at over 60% per year since 2023, and it's actually expected to speed up. These four companies are currently on course to spend over $700 billion on data centers this year alone, compared to the $375 billion they spent in 2025. That's a 90% increase year over year, and it's coming at a huge cost. Amazon's free cash flow went from $18 billion to negative $7.6 billion over the last 12 months. But just a few weeks ago, Amazon increased their CapEx budget for 2026, from $200 billion to $220 billion due to the price of memory chips.

That same week, Alphabet added $15 billion to its CapEx budget while reporting that their free cash flow turned negative for the first time since Google went public in 2004. And Meta's free cash flows are down by 91%. You'd think that spending would slow down once the biggest companies driving it burn through all their cash flows, and then some. But instead, they're borrowing money and issuing new shares. Amazon sold $37 billion worth of bonds in quarter one. Alphabet issued over $65 billion worth of bonds so far this year and announced an $85 billion dollar equity raise in June, the biggest raise in American corporate history. Debt covered 9% of hyperscaler capex in 2024. Today, it covers 32%.

It's important for investors to understand why spending isn't slowing down and why these companies keep investing in AI at all costs. Data center infrastructure is a zero-sum game. There's only so much land and grid-connected power in the first place, and every acre and gigawatt you don't get is one that your competition does. A big new data center can wait anywhere from two to six years for a grid connection. And even when it's all bought up, spending will just shift to the next big bottleneck. Cooling, compute density, network and memory speeds, all of which are zero-sum games too, since every single part of the AI stack is currently supply constrained. So the question isn't whether AI spending will slow down, but where it will shift to next.

And three of the biggest ai chip companies on earth nvidia cerebrus and amd all have very different answers let's start with amd amd's answer is memory an ai model is made up of billions of numbers called parameters and the actual value of each parameter comes from the patterns that it learned during training those values are called weights and each weight takes up about 16 bits or 2 bytes of memory so a model with 400 billion parameters needs around 800 gigabytes of memory a single nvidia h200 gpu has 141 gigabytes of memory so you'd need about six of them just to ask the model a single question today those weights live in memory that sit outside of the processor and the processor spends most of its time waiting for data to arrive that wait time is the bottleneck on august 6th AMD announced plans to acquire a company that fixes this by etching the model directly into the chip.

The company is called Talus, a chip startup founded in 2023, and it etches AI model weights directly into the metal layers of a chip so they don't have to be loaded from memory. The benefits here are huge Total throughput goes way up and latency goes way down when the processor stops having to wait for data to come from memory But the trade is pretty huge too Each chip is permanently dedicated to a single specific AI model exchanging general purpose programmability for maximum efficiency. If the acquisition does go through, AMD will combine these specialized chips with their own Instinct GPUs and rack-scale setups like Helios, so that the GPUs can process prompts and manage dynamic workloads, while these new fixed-weight chips handle the memory-heavy decode phase of inference, kind of like NVIDIA's recent partnership with GroK.

But where GroK's Language Processing Units, or LPUs, have 500 megabytes of ultra-fast on-chip memory, Talos's HC1 chips turn the model directly into hardware.

A weight of 1 is a physically connected wire, and a 0 is a broken path, so the math is automatically executed as data simply flows through the chip without any traditional instruction cycles. And speaking of data, I started getting way more spam phone calls and texts right around the end of the pandemic. When I asked my friends, they said the same thing, so I decided to dig into it and found that hundreds of online data brokers were collecting and selling my personal information—and they might be selling yours too. That's when I joined DeleteMe, the sponsor of this video, and I've been with them ever since.

DeleteMe is a hands-free subscription service that will remove your personal information from those online data brokers. Every quarter you get a privacy report showing you what they've done. I just got my 17th report and DeleteMe reviewed over 60,000 listings from me so far, and they removed my data over 600 separate times—not just my name and address but my wife's and my family's too. But here's the best part: they'll keep scanning these websites even after they remove my information. Talk about a total lifesaver! So if you care about your family's privacy and you like saving money, you can get 20% off any consumer plan with my code SYMBOL20 by going to join delete me dot com slash symbol 20 or with my link in the description. And a big thank you to DeleteMe for supporting the channel and keeping my family's data safe.

Alright, so Talus's chip completely destroys the memory wall, which is the big bottleneck of moving data back and forth between a processor and separate memory. Here's what that means in practice: a typical cloud AI provider can serve a model like Llama 3.1 8B at around 140 tokens per second per user. That same model can hit 17,000 tokens per second on the Talus chip, making it around 120 times faster than a traditional GPU. Responses are generated so fast that entire pages of text appear instantly, rather than streaming in line by line.

But here's the big problem: if the chip is directly tied to a single model and chips take years to design, doesn't that mean that the chip will be obsolete by the time it leaves the production line?

Well, actually, Talus found a way to get around that problem with a pretty clever design. Instead of redesigning the whole chip for every new model, Talus uses a tiered manufacturing process that breaks the chip into more than a hundred physical layers of silicon and metal stacked on top of each other.

Everything underneath the top two layers—the transistors, the power distribution, the math units, and the memory—never has to change. Only the two metal layers at the very top translate a model's specific weights into a physical chip. And since 98% of the chip stays the same between models, they can be made ahead of time and stored until a model is chosen for that chip. Then, once the two layers are designed specifically for a new model, TSMC can pull the base wafer off a shelf, add those two layers, and ship the chip within 60 days.

This two-month turnaround completely shifts the economics around data centers. Instead of buying general purpose GPUs or spending many years and many more billions of dollars designing custom ASICs, they can just buy a cluster of chips hardwired for specific models, run those chips for a year, and then literally swap them out when the next model drops. And not only is it up to 120 times faster than a traditional GPU, it's also roughly 20 times cheaper since it's made on an older 6 nanometer technology.

It has no high bandwidth memory and it uses much simpler chip packaging. Talus also says that their chips cost less than a penny per million tokens to run versus about 3 cents for an Nvidia Blackwell running the same model So when spending shifts from acquiring land buildings and power AMD is betting that it shift into hyper low chips that can be easily swapped out as models keep improving. But while AMD is cutting costs, Cerebris is cutting cables. A modern AI cluster has thousands of separate chips stitched together with network cables and switches. Every time data moves from one chip to another, it costs time and power. And just like with memory, chips can spend more time waiting for the data to arrive over a network than they spend processing it. Think about what this actually means.

But while AMD is cutting costs, Cerebris is cutting cables. A modern AI cluster has thousands of separate chips stitched together with network cables and switches. Every time data moves from one chip to another, it costs time and power. And just like with memory, chips can spend more time waiting for the data to arrive over a network than they spend processing it. Think about what this actually means.

Chipmakers spend tens of thousands of dollars turning a big silicon wafer into dozens of individual chips. Then they spend billions of dollars connecting those chips back together inside data centers. Cerebris is betting their entire company that this approach is wrong. So they never cut the wafer in the first place. Instead, they turn it into one massive chip called the Wafer Scale Engine, or WSE for short. Their current generation is the WSE3, and each side is about 8 inches. That's around 30 times more area than Nvidia's Blackwell B200 chips. If Nvidia's chips are the size of postage stamps, then Cerebris' are the size of dinner plates. But as you know, size doesn't matter. It's all about how you use it.

Cerebris' chips have 4 trillion transistors, that's 19 times more than Nvidia's B200s. But since the chip is around 30 times bigger, that means that Nvidia actually packs over 50% more transistors into the same area, because their chips are made on a more advanced process node by TSMC. The wafer-scale engine also has a whopping 900,000 cores, four times more than Blackwell, and also comes with a whopping 44GB of SRAM, which is the same kind of high-speed memory used in the Grok LPU. Except, you know, 88 times more of it. That memory can move data at 21 petabytes per second, which is about 2,600 times the memory bandwidth of NVIDIA's Blackwell B200s. There are over 18,000 titles on Netflx, and their uncompressed master archive is about 4 petabytes.

That means this Cerebris chip can move Netflx's entire library between cores 5 times every second. That's a huge deal for AI inference performance and it's all because Nvidia has to move their data between chips, across cables and through switches, all of which add extra time to every transfer. While Cerebris simply moves data across one massive chip. No hops, no cables, only compute. Cerebris just reported earnings for the first time since going public and the numbers speak for themselves. The remaining performance obligations hit 25.4 billion dollars. basically their backlog cerebris is guiding for 880 to 890 million dollars in revenue this year so that backlog is worth about 29 times everything they expect to sell in 2026.

They expect to deliver on about 22 of that backlog in the next two years and another 43 in the two years after that. That first chunk works out to about 5.6 billion dollars by the middle of 2028, or roughly 2.8 billion dollars per year, which means their revenue should roughly triple. Revenue from hardware came in at 82 million dollars, which is up 17% from last year. On the official books, hardware sales look like they fell by 23%, but that's only because of a 28 million dollar charge for stock warrants that Cerebris handed to OpenAl. If you remove that charge, their hardware sales grew. This is exactly why I ignore headlines and dig into the numbers myself.

All right, hardware sales are up but only by 17%. On the other hand, revenues from cloud and services hit 127 million dollars, which is up 287% year over year. So the real money for Cerebris is in selling access to their machines, not selling the machines themselves. And their gross margins fell, but mostly because they can't build capacity fast enough to serve all the demand for their chips. So they actually end up renting their own systems back from the cloud companies that bought them to fill the gap. So while AMD and Talus are betting on model-specific chips, Cerebris is betting on wafer-sized ones.

But we can't talk about the future of AI chips without talking about Nvidia. And if you feel I earned it, consider hitting the like button and subscribing to the channel. That really helps and it lets me know to make more content like this. Thanks. Now let's talk about Nvidia. The question isn't whether AI spending will slow down; it's where it will shift to next. Nvidia's answer is to the chips themselves as investable assets. Think about how you'd build a skyscraper: nobody is paying cash. A developer borrows most of the money because the building holds its value, and the lender knows that it can be sold if something goes sideways. That's all an investable asset is: something you can borrow against because everyone agrees it'll still be worth something later.

AI data centers mostly get paid for out of a company's free cash flows, which as I just said earlier is starting to run out. So Nvidia is trying to change where that money comes from altogether. On August 10th, Nvidia announced deals with Apollo, BlackRock, Blackstone, Brookfiel, Goldman Sachs, and KKR to build financing platforms for over half a trillion dollars worth of AI infrastructure.

None of that money is NVIDIA's. It's pension funds, sovereign wealth funds, insurers, the slowest and most conservative money on the planet. Jensen Huang's big idea is that a GPU makes good collateral because someone else will always want it, and NVIDIA's software updates extend its useful lifespan over time.

For example, NVIDIA came out with an open source software package called tensor rt llm which doubled the inference performance of large language models running on h100s that means every h100 already sitting in a data center got roughly twice as good at running large language models overnight for free and that software also works on nvidia's older ampere and ada lovelace chips too and all the way up through blackwell the part that i'm not so sure about is what these older chips will actually sell for on the second-hand market cars lose a lot of value based on their age and how they compare to newer models not just on how fast they drive or their miles per gallon ai chips might depreciate the same way we just don't know yet there are also three big catches to nvidia's deals first none of them are binding at least not yet nobody has actually committed a single dollar second nvidia could quietly be on the hook for the difference an nvidia blog post published the next day said that Nvidia may cover up to 25% of the losses if the equipment ends up being worth less than a loan assumed.

And third, the loans can last longer than the contracts to pay them off. The same day as the announcement, CoreWeave closed a $2.6 billion loan against its GPUs. The loan runs for about five years, while the customer contracts renting those same GPUs only last for three. So if those GPUs don't get rented again 3 years from now, even though they'll be much older, that could cause some serious issues with paying back the 5-year loan. So the big question the market needs to answer is what is the actual useful lifespan of a GPU? Is it 3 years? Is it 5? That determines whether they're worth investing in as an asset class at all. And we might get an answer on August 26th, when Nvidia reports their earnings.

Either way, the biggest case against AI has always been that spending will eventually slow down and when it does every stock that lives or dies by that spending will go down with it but amazon and google both went free cash flow negative and raised capital to spend even more so the question isn't whether ai spending will slow down but where it shifts to next and three of the biggest ai chip companies all have different answers amd is betting it'll shift to hyper-optimized low-cost chips that can easily be swapped out as models keep improving cerebris is betting on wafer-sized chips that cut out rack-level network cables and switches altogether and nvidia is betting that gpus will turn into an asset class of their own because somebody will always want to rent them either way i expect ai spending to keep speeding up until the world runs out of land and power for data centers and after that spending will just shift to filling them money is no longer the constraint here that's why i think investing in ai is still a great way to get rich without getting lucky and if you want to see even more stocks i'm buying to get rich without getting lucky check out this video next either way thanks for watching and until next time this is ticker symbol you my name is alex reminding you that the best investment you can make is in you.