Training Example: Stratechery – Review the Data, Give Your Score & Compare to the Real AI Evaluation

This Week in Stratechery
[H2] 2026.21: The Data Center Veto
(Lexi Critchett/Bloomberg via Getty Images)
Welcome back to This Week in Stratechery!
As a reminder, each week, every Friday, we’re sending out this overview of content in the Stratechery bundle; highlighted links are free for everyone. Additionally, you have complete control over what we send to you. If you don’t want to receive This Week in Stratechery emails (there is no podcast), please uncheck the box in your delivery settings.
On that note, here were a few of our favorites this week.
Data Center Discontent. The impact of AI is, at least for now, being felt digitally: that is where AI is useful, and the more digital a job, the more it is threatened by LLMs. AI, however, depends on data centers in the physical world, and building data centers needs permission. This gives normal people the sort of veto power over AI they didn’t have in the face of globalization; I make the case in Monday’s Update and on Sharp Tech that understanding this dynamic is more important that trying to correct misinformation, which is a symptom, not a cause, of data center opposition. — Ben Thompson

Continue reading

Agent Economics. What will the internet look like when ad-supported models are rendered obsolete by shifting user behavior and the rise of agentic web traffic? Ben considered this question last summer with The Agentic Web and Original Sin, and I was surprised to learn this week that Parag Agarwal, former CEO of Twitter, is now focused on devising solutions for exactly this reality. This week’s Stratechery Interview with Agarwal dives deep into the economics of content on the Internet, why ads make sense for humans, and why incentivizing content for agents will be different, and how Agarwal and Parallel are trying to solve them. I learned a ton from this interview, and I bet you will, too — and don’t worry, we did get a few bonus questions on the ride at Twitter.  — Andrew Sharp
Never Count Out the Slime Mold. Wednesday’s Daily Update on Google I/O reminded me of an iconic leaked memo about the ungovernable and poorly coordinated mold in Mountain View, as the company seems to be throwing 10 different types of AI spaghetti at the wall to see what sticks. Then again, Google is now a nearly $5 trillion company and its transformer architecture supercharged the AI era. That second part is why, when Ben highlights a DeepMind approach to building AGI that’s distinct from the approaches at OpenAI and Anthropic, I’m compelled to both pay attention, and remember: for all of Google’s faults and misses, they do in fact have plenty of historic hits.  — AS
[H3] Stratechery Articles and Updates
Data Center Discontent, Understanding the Opposition, Fixing the Problem — There are understandable reasons for people to oppose data centers; the only solution that will work is simply paying them off.
Google I/O, World Models, I/O Spaghetti — Google I/O put AI everywhere, for better and for worse. Meanwhile, is DeepMind aligned with Google’s business objectives?
An Interview with Parallel Founder Parag Agarwal About Valuing Content on the Agentic Web — An interview with Parallel founder Parag Agarwal about valuing content and incentivizing its creation in a world of agents (plus questions about Twitter).
[H3] Dithering with Ben Thompson and Daring Fireball’s John Gruber
Data Center Unpopularity
Google Being Google
[H3] Asianometry with Jon Yu
The Little Vertical Laser That Everyone Uses
Intel’s 30 Years in Costa Rica
[H3] Sharp China with Andrew Sharp and Sinocism’s Bill Bishop
Constructing US-China Stability; Trump’s Taiwan Comments and More Summit Takeaways; Putin in China
[H3] Greatest of All Talk
Wemby, Harper and an Instant Classic from the Spurs in Game 1 vs. OKC
A Note on the Future of GOAT and An Emergency Top Five
[H3] Sharp Tech with Andrew Sharp and Ben Thompson
Much Ado About Data Centers, What Tech Gets Wrong About Its Critics, Q&A on SpaceX, Chinese AI, Elon Musk
This week’s Stratechery video is on The Inference Shift.
[H2] The Inference Shift
Monday, May 11, 2026
Listen to Podcast
Watch on YouTube
Listen to this post:
Log in to listen
If you were looking for the ideal time to IPO, being a chip company in May 2026 is hard to beat. Reuters reported over the weekend:
Cerebras Systems is set to raise the size and price of its initial public offering as soon as Monday, as demand for the artificial intelligence chipmaker’s shares continues to climb, two people familiar with the matter told Reuters on Sunday. The company is considering a new IPO price range of $150-$160 a share, up from $115-$125 a share, and raising the number of shares marketed to 30 million from 28 million, said the sources, who asked not to be identified because the information isn’t public yet.
The fundamental driver of the ongoing surge in semiconductor stocks is, of course, AI, particularly the realization that agents are going to need a lot of compute. What Cerebras represents, however, is something broader: while the compute story for AI has been largely about GPUs, particularly from Nvidia, the future is going to look increasingly heterogeneous.
[H3] The GPU Era
The story of how Graphics Processing Units became the center of AI is a well-trodden one, but in brief:
Just as drawing pixels on a computer screen was a parallel process, which meant there was a direct connection between the number of processing units and graphics speed, making AI-related calculations was a parallel process, which meant there was a direct connection between the number of processing units and calculation speed.
Nvidia enabled this dual-usage by making its graphics processors programmable, and created an entire software ecosystem called CUDA to make this programming accessible.
The big difference between graphics and AI has been the size of the problem being solved — models are a lot bigger than video game textures — which has led to a dramatic expansion in high-bandwidth memory (HBM) per GPU, and dramatic innovations in terms of chip-to-chip networking to allow multiple chips to work together as one addressable system. Nvidia has been the leader in both.
The number one use case for GPUs has been training, which stresses the third point in particular. While the calculations within each training step are massively parallel, the steps themselves are serial: every GPU has to share its results with every other GPU before the next step can begin. This is why a trillion-parameter model needs to fit in the aggregate memory of tens of thousands of GPUs that can communicate as one system. Nvidia dominates both problem spaces, first by securing HBM ahead of the rest of the industry, and second thanks to its investments in networking.
Of course training isn’t the only AI workload: the other is inference. Inference has three main parts:
Prefill encodes everything the LLM needs to know into an understandable state; this is highly parallelizable and compute matters.
The first part of decode entails reading the KV cache — which stores context, including the output of the prefill step — to make an attention calculation. This is a serial step where bandwidth matters, but the memory requirements are variable and increasingly large.
The second part of decode is the feed-forward computation over the model weights; this is also a serial step where bandwidth matters, and the memory requirements are defined by the size of the model.
The two decode steps alternate for every layer of the model (they’re interleaved, not in sequence), which is to say that decode is serial and memory-bandwidth bound. For every token generated, two distinct memory pools must be read: the KV cache, which stores context and grows with each token, and the model weights themselves. Both must be read in full to produce a single output token.
GPUs handle all three needs: high compute for prefill, abundant HBM for KV cache and model weights, and chip-to-chip networking to pool memory across multiple chips when a single GPU isn’t enough. In other words, what works for training works for inference — look no further than the deal SpaceX made with Anthropic. From Anthropic’s blog:
We’ve signed an agreement with SpaceX to use all of the compute capacity at their Colossus 1 data center. This gives us access to more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month. This additional capacity will directly improve capacity for Claude Pro and Claude Max subscribers.
SpaceX retains Colossus 2 — presumably for both training of future models and inference of existing ones — and can afford to do both in the same data center precisely because xAI’s models aren’t getting much usage; more pertinently to this piece, they can do both in the same data center because both training and inference can be done on GPUs. Indeed, the GPUs Anthropic is contracting for at Colossus 1 were originally used for training as well; the fact that GPUs are so flexible is a big advantage.
[H3] Understanding Cerebras
Cerebras makes something completely different. While a silicon wafer has a diameter of 300mm, the “reticle limit” — the maximum area that a lithography tool can expose on that wafer — is around 26mm x 33mm. This is the effective size limit for chips; going beyond that entails linking two separate chips together over a chip-to-chip interposer, which is exactly what Nvidia has done with the B200. Cerebras, on the other hand, has invented a way to lay down wiring across the so-called “scribe lines” that are the boundary between reticle exposures, making the entire wafer into a single chip with no need for relatively slow chip-to-chip linkages.
The net result is a chip with a lot of compute and a lot of SRAM that is blisteringly fast to access. To put it in numbers, the WSE-3 (Cerebras’ latest chip) has 44GB of on-chip SRAM at 21 PB/s of bandwidth; an H100 has 80GB of HBM at 3.35 TB/s. In other words, the WSE-3 has just over half the memory of an H100, but 6,000 times the memory bandwidth.
The reason to compare the WSE-3 to an H100 is that the H100 is the chip most used for inference — and inference is clearly what Cerebras is most well-suited for. You can use Cerebras chips for training, but the chip-to-chip networking story isn’t very compelling, which is to say that all of that compute and on-chip memory is mostly just sitting around; what is much more interesting is the idea of getting a stream of tokens at dramatically faster speed than you can from a GPU.
Note, however, that the limitation in terms of training also potentially applies in terms of inference: as long as everything fits in on-chip memory Cerebras’ speed is an incredible experience; the moment you need more memory, whether that be for a larger model or, more likely, a larger KV cache, then Cerebras doesn’t make much sense, particularly given the price. That whole-wafer-as-chip technique means high yields are a massive challenge, which hugely drives up costs.
At the same time, I do think there will be a market for Cerebras-style chips: right now the company is highlighting the usefulness of speed for coding — reasoning means a lot of tokens, which means that dramatically scaling up tokens-per-second equals faster thinking — but I think this is a temporary use case, for reasons I’ll explain in a bit. What does matter is how long humans are waiting for an answer, and as products like AI wearables become more of a thing, the speed of interaction, particularly for voice — which will be a function of token generation speed — will have a tangible effect on the user experience.
[H3] Agentic Inference
I have previously made the case, including in Agents Over Bubbles, that we have gone through three inflection points in the LLM era:
ChatGPT demonstrated the utility of token prediction.
o1 introduced the idea of reasoning, where more tokens meant better answers.
Opus 4.5 and Claude Code introduced the first usable agents, which could actually accomplish tasks, using a combination of reasoning models and a harness that utilized tools, verified work, etc.
All of this falls under the banner of “inference”, but I think it will be increasingly clear that there is a difference between providing an answer — what I will call “answer inference” — and doing a task — what I will call “agentic inference.” Cerebras’ target market is “answer inference”; in the long run, I think the architecture for “agentic inference” will look a lot different, not just from Cerebras’ approach, but from the GPU approach as well.
I mentioned above that fast inference for coding is a temporary use case. Specifically, coding with LLMs requires a human in the loop. It’s the human that defines what is to be coded, checks the work, commits the pull request, etc.; it’s not hard to envision a future, however, where all of this is completely handled by machines. This will apply to agentic work broadly: the true power of agents will not be that they do work for humans, but rather that they do work without human involvement at all.
This, by extension, will mean that the likely best approach to solving agentic inference will look a lot different than answer inference. The most important aspect for answer inference is token speed; the most important aspect for agentic inference, however, is memory. Agents need context, state, and history. Some of that will live as active KV cache; some will live in host memory or SSDs; much of it will live in databases, logs, embeddings, and object stores. The important point is that agentic inference will be less about GPUs answering a question and more about the memory hierarchy wrapped around a model.
Critically, this articulation of an agentic-specific memory hierarchy implies a necessary trade-off of speed for capacity. Here’s the thing, though: lower speed isn’t nearly as important a consideration if there isn’t a human in the loop. If an agent is waiting around for a job that is being run overnight, the agent doesn’t know or care about the user experience impact; what is most important is being able to accomplish a task, and if entirely new approaches to memory make that possible, then delays are fine.
Meanwhile, if delays are fine, then all of the focus on pure compute power and high-bandwidth memory seems out of place: if latency isn’t the top priority, then slower and cheaper memory — like traditional DRAM, for example — makes a lot more sense. And if the entire system is mostly waiting on memory, then chips don’t need to be as fast as the cutting edge either. This represents a profound shift in future architectures, but it also doesn’t mean that current architectures are going away:
Training will continue to matter, and Nvidia’s current architecture, including high-speed compute, large amounts of high-bandwidth memory, and high-speed networking, will likely continue to dominate.
Answer inference will be a meaningful market, albeit a relatively small one, and speed from chips like Cerebras or Groq (I explained how N

[H1]
[IMG: Stratechery Plus]
Subscribe to Stratechery Plus for full access.
Already subscribed?
Log in
$15 / month or $150 / year
Subscribe to Stratechery Plus
With Stratechery Plus you get access to the subscriber-only Stratechery Update and Stratechery Interviews, and the Sharp Tech, Sharp China, Dithering, Greatest of All Talk, and Asianometry podcasts.
[IMG: The Stratechery Update]
Stratechery UpdateSubstantial analysis of the news of the day delivered via three weekly emails or podcasts.
[IMG: Stratechery Interviews]
Stratechery InterviewsInterviews with leading public CEOs, private company founders, and discussions with fellow analysts.
[IMG: Dithering]
DitheringA twice-weekly podcast from John Gruber and myself: 15 minutes an episode, not a minute less, not a minute more.
[IMG: Sharp Tech with Ben Thompson]
Sharp TechAndrew Sharp and myself discuss how technology works and the ways it impacts our lives.
[IMG: Sharp China with Sinocism]
Sharp ChinaA weekly podcast from Andrew Sharp and Sinocism’s Bill Bishop about understanding China and how China impacts the world.
[IMG: The Greatest Of All Talk]
Greatest Of All TalkA twice-weekly podcast from Andrew Sharp and Ben Golliver about the NBA, life, and national parks.
AsianometryAudio and transcripts of the Asianometry YouTube channel, the best source for learning about how tech works.
Sharp TextSharp Text is an extension of GOAT, Sharp Tech and Sharp China, where Andrew writes about basketball, technology, and US-China relations with weekly posts.
Stratechery Updates are also available via SMS, RSS, or on this site. Please see the Stratechery Update Schedule for more details about delivery times and planned days-off. Please note that all subscriptions auto-renew monthly/annually (but can be cancelled at any time). If you are interested in ordering and managing multiple subscriptions for your team or company, please fill in the form here.
[H4] Frequently-Asked Questions
How do I subscribe to the Stratechery Podcast?
Once you are subscribed, please visit your Delivery Preferences where you will find easy-to-follow instructions for adding Stratechery Podcasts to your favorite podcast player.
Can I read Stratechery via RSS?
Yes! Create a Stratechery Passport account, go to Delivery Preferences, and add your personalized RSS feed. Free accounts will have access to Weekly Articles, while subscribers will have access to the Daily Update as well.
Can I share a Stratechery Update subscription with a friend?
No, the Stratechery Update and Stratechery Podcast are intended for one subscriber only. Sharing emails, using shared inboxes, or sharing RSS feeds is a violation of Stratechery’s Terms of Service, and your account may be suspended or your RSS feed reset. Of course occasional forwarding of the Stratechery Update to interested friends or colleagues is totally fine.
Can I buy a subscription for my team?
Yes! You can purchase a team subscription here.
Can I switch to an annual plan?
Yes! Just go to your account page, choose the ‘Subscriptions’ tab, and click the Annual upgrade button. You will be charged immediately, with a prorated discount applied for the remainder of your current monthly plan.
Do you offer a student discount?
Stratechery is purposely kept at a low price — thousands of dollars less than other analyst reports or newsletters — to ensure it is accessible to everyone, including students.
Can you create a custom invoice that meets my government/company requirements?
I am happy to create an invoice to your specification for annual subscribers; however, it is simply not viable for me to offer this service to monthly subscribers. Therefore, if you need a custom invoice please subscribe or switch to an annual subscription and contact Stratechery.
June 1, 2021 Update: We are hoping to add native support for custom invoices to Passport; you can subscribe to Passport Updates to be notified when it is available.
Can I give a subscription as a gift?
Yes! To send a gift visit the gifts page.

[H2] An Interview with Parallel Founder Parag Agarwal About Valuing Content on the Agentic Web
Thursday, May 21, 2026
Listen to Podcast
An interview with Parallel founder Parag Agarwal about valuing content and incentivizing its creation in a world of agents (plus questions about Twitter).
[H1]
[IMG: Stratechery Plus]
Subscribe to Stratechery Plus for full access.
Already subscribed?
Log in
$15 / month or $150 / year
Subscribe to Stratechery Plus
With Stratechery Plus you get access to the subscriber-only Stratechery Update and Stratechery Interviews, and the Sharp Tech, Sharp China, Dithering, Greatest of All Talk, and Asianometry podcasts.
[IMG: The Stratechery Update]
Stratechery UpdateSubstantial analysis of the news of the day delivered via three weekly emails or podcasts.
[IMG: Stratechery Interviews]
Stratechery InterviewsInterviews with leading public CEOs, private company founders, and discussions with fellow analysts.
[IMG: Dithering]
DitheringA twice-weekly podcast from John Gruber and myself: 15 minutes an episode, not a minute less, not a minute more.
[IMG: Sharp Tech with Ben Thompson]
Sharp TechAndrew Sharp and myself discuss how technology works and the ways it impacts our lives.
[IMG: Sharp China with Sinocism]
Sharp ChinaA weekly podcast from Andrew Sharp and Sinocism’s Bill Bishop about understanding China and how China impacts the world.
[IMG: The Greatest Of All Talk]
Greatest Of All TalkA twice-weekly podcast from Andrew Sharp and Ben Golliver about the NBA, life, and national parks.
AsianometryAudio and transcripts of the Asianometry YouTube channel, the best source for learning about how tech works.
Sharp TextSharp Text is an extension of GOAT, Sharp Tech and Sharp China, where Andrew writes about basketball, technology, and US-China relations with weekly posts.
Stratechery Updates are also available via SMS, RSS, or on this site. Please see the Stratechery Update Schedule for more details about delivery times and planned days-off. Please note that all subscriptions auto-renew monthly/annually (but can be cancelled at any time). If you are interested in ordering and managing multiple subscriptions for your team or company, please fill in the form here.
[H4] Frequently-Asked Questions
How do I subscribe to the Stratechery Podcast?
Once you are subscribed, please visit your Delivery Preferences where you will find easy-to-follow instructions for adding Stratechery Podcasts to your favorite podcast player.
Can I read Stratechery via RSS?
Yes! Create a Stratechery Passport account, go to Delivery Preferences, and add your personalized RSS feed. Free accounts will have access to Weekly Articles, while subscribers will have access to the Daily Update as well.
Can I share a Stratechery Update subscription with a friend?
No, the Stratechery Update and Stratechery Podcast are intended for one subscriber only. Sharing emails, using shared inboxes, or sharing RSS feeds is a violation of Stratechery’s Terms of Service, and your account may be suspended or your RSS feed reset. Of course occasional forwarding of the Stratechery Update to interested friends or colleagues is totally fine.
Can I buy a subscription for my team?
Yes! You can purchase a team subscription here.
Can I switch to an annual plan?
Yes! Just go to your account page, choose the ‘Subscriptions’ tab, and click the Annual upgrade button. You will be charged immediately, with a prorated discount applied for the remainder of your current monthly plan.
Do you offer a student discount?
Stratechery is purposely kept at a low price — thousands of dollars less than other analyst reports or newsletters — to ensure it is accessible to everyone, including students.
Can you create a custom invoice that meets my government/company requirements?
I am happy to create an invoice to your specification for annual subscribers; however, it is simply not viable for me to offer this service to monthly subscribers. Therefore, if you need a custom invoice please subscribe or switch to an annual subscription and contact Stratechery.
June 1, 2021 Update: We are hoping to add native support for custom invoices to Passport; you can subscribe to Passport Updates to be notified when it is available.
Can I give a subscription as a gift?
Yes! To send a gift visit the gifts page.
[H3] Share
Share on Facebook (Opens in new window)
Facebook
Share on X (Opens in new window)
X
Share on LinkedIn (Opens in new window)
LinkedIn
Email a link to a friend (Opens in new window)
Email
←Google I/O, World Models, I/O Spaghetti
2026.21: The Data Center Veto→