Rihard Jarc

I just published my $GOOGL TPU deep dive. 1. Performance numbers TPU vs $NVDA (Former $GOOGL employees, customers, and competitors weigh in) 2. Why the TPU is $GOOGL's most important moat for the next 10 years 3. Why $GOOGL will save +$40B in CapEx next year because of the TPU 4. Gemini 3 and the changes that are coming to the industry and much more. There aren't many comprehensive articles out there covering this topic, so I hope you enjoy reading it as much as I enjoyed researching it! uncoveralpha.com/p/the-chip-mad…

$AMZN chip business just keeps getting wins. Altman on Trainium: “We are quite excited to get these models running on Trainium” AWS CEO: “…over time, more and more of it will be on Trainium."

@DicksonPau Matthew Berman's (timestamp 06:57): youtube.com/watch?v=bNdiBw…

@RihardJarc This is from which podcast?

I still can't get past the recent comment made by Thomas Kurian, the CEO of $GOOGL's GCP, regarding compute: "I think for the next 10 years there will always be more demand than supply". This is not a comment from OAI or Anthropic, who, in some sense, need to use marketing to raise new funds; it's from $GOOGL, which, in many ways, benefits more if the pace of AI is slower.

@98_tennis It does help AWS a lot if OAI continues to ship SOTA models. Also enterprises want an orchestration layer where they can switch between model providers for different work use-cases, as we will have more and more specializations.

@RihardJarc so does the updated msft and OpenAI deal benefit amzn the most ?

Reading between the lines of the $MSFT - OpenAI amended deal and the recent comments out of OAI you can really see how important $AMZN AWS is and how much demand Bedrock has.

I just published my Q1 alternative data report covering the cloud providers $AMZN, $GOOGL, and $MSFT: - Data showing significant growth - Expecting higher cloud growth than the consensus from one particular provider - Surprising data on $MSFT Copilot uncoveralpha.com/p/q1-2026-chan…

$META just partnered with $AMZN to start deploying tens of millions of their Graviton CPU cores. $META is now the largest Graviton customer in the world. Another sign that the CPU shortage is real, and secondly, that both $AMZN and $GOOGL are chip companies.

I have talked & written about the importance of $INTC in addressing 3 key bottlenecks: CPUs, advanced packaging, and fab capacity. On this earnings call, we got confirmation of $INTC executing on all three areas. For the future, aside from the obvious CPU shortage, it is a great sign to see $INTC lift its 2026 CapEx. $INTC often talked about that being the key sign for 14A adoption, and you are getting it now.

@johnnyjiayu $META cut a lot even before in the last years, so this is more about them making room for CapEx IMO.

@RihardJarc How much is this previous to over-hiring? And how much is it a switch from labour costs to AI compute?

$META cutting 10% of the workforce (8.000 roles) and eliminating 6.000 open positions. This will happen across the board in the tech sector and more broadly in knowledge work, Zuck is as always just ahead of the pack. Companies are switch labour costs for AI compute.

@silverhawk_ny I don’t think they are done for the year even.

@RihardJarc 10% seems not that much , I thought it will be 20% level

Agree on this and actually wrote about this as I believe we are coming to the end of this subsidized era this year. At the same time I am not convinced that even if these labs raised priced substantially, say to their costs that demand would slow down that significantly. uncoveralpha.com/p/the-era-of-s…

I find this conversation foreign - along with the argument that we are "data center constrained" or "energy constrained." Historically, in markets - price is the leveler of supply and demand. If you have a constraint, you price higher - you don't have "surplus demand." But in this market, VC$$$ act as subsidies (as they did in consumer internet). Everyone believes if they have high growth they get unlimited VC$$$. The biggest fear becomes losing market share. So "growth at all costs" becomes the game on the field. With that reality, you are always going to have some constraint, because you are "knowingly" choosing pricing that is out of whack with balancing supply/demand. It will continue until the major players feel they are forced to reconcile unit economics and profitability (as eventually happened in ride sharing when Lyft went public). Until then, you by definition have constraints. We can’t disentangle true demand from subsidized demand yet. Some of the incremental demand is being engineered by the excessive VC$$$ forced into the system, and the competitive dynamic between two companies that are losing massive amounts of money. No one can argue they aren't losing tons of money. Amazon and Uber maxed out around $2B a year. These companies could lose $10B on more in 2026.

Dylan on why Anthropic leading on models doesn't mean OpenAI loses. Even the tier 2 and 3 labs will sell out of tokens: "Everyone's like Anthropic won. They had Mythos, Opus 4.7 and their revenue is adding $10 billion a month all before OpenAI's alleged Spud release. So clearly Anthropic is in the lead and OpenAI is cooked. Anthropic has such bounds on compute, they can only grow it so fast. Dario used to gloat about how OpenAI was being too aggressive on compute and Anthropic was more sensible in their scaling. Now Anthropic is like I wish we had a lot more compute. OpenAI just raised a ton of money to get incremental compute in addition to the irresponsible levels of compute that they were buying. By the end of the year, let's say a 4.6 Opus tier model, the economy would spend $100 billion on. Anthropic won't have enough compute to do that. And presumably OpenAI and Google will hit that tier soon. Anthropic may get to charge 70%+ gross margins but if OpenAI hits it next, they charge 50% gross margins. They still get all of this incremental demand. There is such demand for these tokens and such limitations on compute. It's pretty clear even the tier 2 or tier 3 lab are going to be sold out of tokens. The economic value that the best model can deliver is growing faster than our ability to actually serve those tokens to people via the infrastructure."

If you found this insightful, make sure to subscribe to my newsletter, as I will be publishing my alternative data report with more insights like this one tomorrow, before big tech earnings start next week.

Very insightful interview with a $AMZN AWS employee explaining how the demand dynamic of GPUs and power looks like today: 1. According to him, in Q1 for AWS, there was around 20% deficit of what their needs were for GPUs, meaning they couldn't get enough GPUs. The deficit used to be smaller than 20% in the past (even in Q3 2025, when they had the highest demand spikes). 2. He mentions the problem of many clients wanting to upgrade to $NVDA's B300, and that they didn't have that problem with B200. Now, a lot of B200 orders have moved to B300, and he thinks a similar situation will happen with Rubin. 3. In his view, in terms of performance, he thinks $NVDA is best at inference as they have 35 times better throughput per megawatt for inference than any other GPU in the market. For training, $NVDA is still one of the best, but he does mention that a lot of companies, $AMZN, $GOOGL have really good training chips, but they are not selling them directly. 4. He talks about other constraints outside of GPUs, like power, that are tapering down growth. He would probably need 4GW of power quickly, but he can probably get half of that. It is also hard to get power, as you are competing with other hyperscalers that also want that much power. 5. In his view, the GPU cycle is 3-5 years. He also mentions the problem of clients now wanting to switch to the newest models sooner. He also explained that when a newer model comes out, the older racks are probably about 25% cheaper, and towards the end of their life, they lose most of their value. My comment: The compute shortage is real and getting more severe. found on @AlphaSenseInc

@98_tennis I have

@RihardJarc You load it then

The market is overly pesimistic on $MSFT. Usage of $MSFT tools is not dropping and Azure is one of the key backbones of the AI economy growing 40% YoY. From the data I am seeing enterprises still love $MSFT especially when it comes to cyber which is a big emphasis now.

@tengyanAI yeah, although OpenAI only has a game plan for owning data centers and hardware, while $GOOGL is already there...

@RihardJarc same game plan as openAI - will be a race to see who dominates in the coming years

The chip company $GOOGL just announced a new version of TPUs. TPU 8t for training and TPU 8i for inference. - TPU 8t 124% better performance per watt vs last generation - TPU 8i 117% better performance per watt vs last generation - TPU 8i is also 80% better performance per dollar Those numbers are really strong considering TPUv7 was already a strong chip for $GOOGL for the Gemini 3 family models, and considering that all signs point to Anthropic's Mythos being trained on TPUv7. $GOOGL owning the whole AI stack from hardware, data center, frontier AI model to distribution will be a significant advantage in the coming years.