# AI Infrastructure, Compute Scarcity, and Geopolitical Shifts

**Podcast:** The AI Daily Brief (Formerly The AI Breakdown): Artificial Intelligence News and Analysis
**Published:** 2026-04-28

## Transcript

Today on the AI Daily Brief, DeepSeek releases their latest model and we're discussing what it has to do with the White House invoking the Defense Production Act around the U.S.
electric grid.
Before that, in the headlines, the AI trade is back and Google plans to invest up to $40 billion in Anthropic.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, Assembly, Blitzy, Granola, and Superintelligent.
To get an ad-free version of the show, go to patreon.com slash AIDailyBrief, or you can subscribe on Apple Podcasts.
If you want to learn more about sponsoring the show, send us a note at sponsors at AIDailyBrief.ai.
And before we get into the show, I did also want to flag that we are now live with our latest free self-directed training program, Agent OS.
This is a program to help you build your personal agentic operating system.
It's similar in some ways to Claw Camp, but this is going to be platform neutral.
and designed to create a system that can evolve over time as the technology changes as well.
Now, if you want to learn more about this, on Saturday, I did an Operator's Bonus episode with Nufar Gaspar, who created this program.
So go check that out, and then you can sign up for AgentOS.
You can find a link from aidailybrief.ai, and join the more than a thousand people who have already signed up for this thing in the first day or so.
With that out of the way, let's dive in.
Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes.
Late last Friday saw just an absolute slew of stories, including that Google has expanded their relationship with Anthropic with another huge investment.
Google and Anthro confirmed a $40 billion investment deal to the press on Friday.
The deal will consist of $10 billion up front and a further $30 billion based on Anthropic hitting undisclosed commercial milestones.
Google had previously invested in Anthropic in late 2023 and early 2025, with the investment totaling $3 billion.
They also signed an agreement to supply 5 gigawatts of compute capacity last month in collaboration with Broadcom.
Those chips are expected to start coming online next year.
Google's investment seems similar to the deal that was struck with Amazon earlier in the month.
Amazon committed $5 billion up front, with a further $20 billion contingent upon, once again, Anthropic hitting certain commercial milestones.
That deal also saw Anthropic commit to spending $100 billion with AWS over the coming decade.
and included 5 gigawatts of supply to come online by the end of this year.
TLDR to many of the Amazon deal looked like Anthropic trading equity for compute.
And although we don't know all the contours of the Google deal, it seems to be of a similar nature.
Citrini analyst Ju Khan republished an excerpt from a Mirai Securities note regarding the Amazon deal, calling it one of the most interesting takes that he had read.
The note stated, all of Anthropic's unusual moves around the GPT 5.5 launch ultimately converge on a single conclusion.
In order to secure compute, Anthropic must bind itself far more deeply and far more dependently to those who possess these physical resources.
The note remarked that each gigawatt of capacity is roughly equivalent to a full-scale nuclear reactor.
Further, Microsoft's entire global data center footprint in 2024 was around 6 gigawatts, with Mireille commenting, This means Anthropic alone is locking in incremental capacity for AI training and serving that rivals the entirety of Microsoft's historical physical infrastructure.
Coupled with Anthropic's announcement of having reached 30 billion ARR, the market is reading this deal as Anthropic pre-signing Amazon's invoice in order to keep its growth growing.
Now, Anthropic, of course, is not alone in adding a ton of capacity.
Last week, OpenAI said that they expect to build out a total of 30 gigawatts of capacity by 2030, with 8 gigawatts already quote-unquote identified.
The difference is OpenAI seems to be partnering with a consortium of Oracle, data center developers, and smaller NeoClouds alongside the established cloud giants.
Amazon invested $50 billion in OpenAI in February, and Microsoft still holds their estimated 50% stake in the company.
Again, from Mirai Securities, what matters is that the structure of this deal is more favorable to Amazon than to Anthropic.
Amazon has already invested up to $50 billion in OpenAI as well.
In other words, the more fiercely OpenAI and Anthropic compete to eat each other's lunch, the more Amazon benefits simultaneously along three axes.
Cloud usage fees from both, adoption rates for its in-house silicon, and visibility on the recovery of data center capex.
This is structurally almost identical to the valuation premium Google historically enjoyed under the full stack player framing.
Now, Mireille's assessment is that the market isn't close to pricing in how the spoils of AI competition will flow to the cloud giants, with Amazon still trading near a 10-year low in terms of revenue multiple.
While some read this as Google giving up on Gemini, others viewed it as a hedge, or even just the development of an entire new business line and new play in the AI space.
What little we know of the deal terms suggests that Google is extracting a healthy premium for their compute.
If the deal terms are on Anthropic's previous $350 billion valuation from their February round, Google could be getting more than a 50% discount to the $800 billion valuation that shares have been flying around on secondary markets at.
Their ownership stake is also beginning to stack up.
The New York Times previously reported that Google owned 14% of Anthropic as of last March, and if they make this full investment, their stake could be heading north of 20%.
Now, this idea that we're undercounting the success of the infrastructure providers is a theme that's starting to grow.
Bloomberg Steve Howe writes, Has Amazon's moment finally arrived?
The speed and degree of the success of Anthropic's singular bet on AGI via coding are unexpected and unprecedented.
Both Anthropic and Google failed to plan and secure a sufficient compute.
In what seems to be years of critical shortage, those that do have the compute to give out find themselves having a lot more strategic leverage.
Contrary to conventional wisdom, I think most suppliers in the AI value chain had all vividly remembered and overlearned the cautionary tales of past overbuilds, especially leading up to the dot-com bubble.
The cautiousness and reluctance of the Korean memory makers to expand capacity are a good example.
Suddenly, almost everyone is caught surprised by the sharp rise of agentic AI demand that could outstrip the supply of AI compute by potentially several orders of magnitude, causing maxed-out productive capacity everywhere in the supply chain, and rationing and price hikes on the end users, at least in the US.
This, I think, is going to be one of the loudest stories banging about the rest of 2026.
Now, speaking of, Meta has also signed a multi-billion dollar deal to rent AI chips from Amazon, although it comes with an interesting twist.
Rather than signing up for Amazon's Tranium ASICs, Meta is renting Amazon's Graviton 5 CPUs.
The CPUs are optimized for agentic workloads, and there is an increasing conversation around whether the CPU architecture could prove to be more efficient than GPUs when it comes to actually running agents.
Given that Meta's entire strategy is centered on delivering consumer agents, the commitment to agent-specific architecture makes a lot of sense.
At the same time, this could also be Meta just placing as many bets as possible and snapping up every chip they can get their hands on.
Relative to their size, Meta has the largest AI build-out, forecast at up to $135 billion this year.
They already rent GPU clusters from AWS and have multi-billion dollar deals with NVIDIA, AMD, Google, CoreWeave, and Nebius.
Announcing the deal, Meta wrote, Now alongside all of this, it appears the AI trade is back as the hyperscalers lead the stock market to all-time highs.
The S&P 500 is up 12.5% over the past month.
completely erasing the drawdown from the beginning of the Iran war.
The Wall Street Journal framed the recovery as all about AI.
They noted that 118 stocks have fallen more than 10% since the war began, and in contrast, the 82 stocks that are up more than 10% are almost exclusively related to AI.
Without the MAG7 tech stocks plus Broadcom, the S&P 500 is actually down year-to-date, meaning the AI industry is once again putting the market on its back.
The journal did note signs of froth in the market from Allbirds' questionable AI pivot, to the massive premium placed on pre-IPO AI stocks.
Still, the fundamental driver is massive in renewed commitments to AI spend.
This week even saw dot-com darling Cisco reach a new all-time high on the back of data center spending, taking 26 years to claw back to their 2000 peak.
And while the WSJ's narrative still focuses on questioning the AI bubble, some analysts are insisting that this time is different.
Corey Acri, a senior semiconductor analyst at Benchmark, said, We just had a return to optimism around the AI trade.
I think the optimism around the demand is correct.
The demand, spending, the CapEx budgets are real.
Ultimately, it's kind of difficult to argue with the numbers.
CapEx commitments are firming up and could actually end the year higher than forecast.
On the back of increasing demand, NVIDIA closed the week at a new all-time high, becoming the world's first $5 trillion company just nine months after they became the first $4 trillion.
To quote OpenAI's Rune, not enough people are emotionally prepared for if it's not a bubble.
For now though, that's going to do it for the headlines.
Next up, the main episode.
One of the trends that I follow most closely when it comes to AI is around voice.
Today's episode is brought to you by Assembly AI, the best way to build voice AI apps.
The company has been moving with extreme velocity lately, shipping major improvements to their speech-to-text models that go way beyond just better transcription.
Specifically, they're getting to an accuracy level that can reliably capture the type of things that used to break every other speech-to-text model.
Think credit card numbers read aloud, email addresses spelled out, complex medical terminology, financial figures.
All of these things, in other words, that it really matters to get right.
So for anyone who's building in fintech, healthcare, sales intelligence, customer support, getting those things wrong isn't just annoying, it's a liability.
Their speech understanding models are also really good at things like identifying speakers, surfacing key moments, and uncovering insights from voice data.
And all of that happens in a single API call.
The proof is in the pudding, and Assembly powers some of the top voice AI products in the market today, like Granola, Dovetail, and Ashby.
Getting started is free.
Head to assemblyai.com slash brief to test it live and get $50 in free credits.
No contract, no upfront commitments.
That's assemblyai.com slash brief.
With the emergence of AI code generation in 2022, NVIDIA master inventor and Harvard engineer Sid Paresci took a contrarian stance.
Inference time compute and agent orchestration, not pre-training, would be the key to unlocking high-quality AI-driven software development in the enterprise.
He believed the real breakthrough wasn't in how fast AI could generate code, but in how deeply it could reason to build enterprise-grade applications.
While the rest of the world focused on co-pilots, he architected something fundamentally different.
Blitzy, the first autonomous software development platform leveraging thousands of agents that is purpose-built for enterprise-scale codebases.
Fortune 500 leaders are unlocking 5x engineering velocity and delivering months of engineering work in a matter of days with Blitzy.
Transform the way you develop software.
Discover how at Blitzy.com.
That's B-L-I-T-Z-Y dot com.
Today's episode is brought to you by Granola.
Granola is the AI notepad for people in back-to-back meetings.
You've probably heard people raving about Granola.
It's just one of those products that people love to talk about.
I myself have been using Granola for well over a year now, and honestly, it's one of the tools that changed the way I work.
Granola takes meeting notes for you without any intrusive bots joining your calls.
During or after the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes which are pre-made prompts.
Once you try it on a first meeting, it's hard to go without.
Head to granola.ai.ai daily and use code AIDaily.
New users get 100% off for the first three months.
Again, that's granola.ai.ai daily.
It is a truth universally acknowledged that if your enterprise AI strategy is trying to buy the right AI tools, you don't have an enterprise AI strategy.
Turns out that AI adoption is complex.
It involves not only use cases, but systems integration, data foundations, outcome tracking, people and skills, and governance.
My company, Superintelligent, provides voice agent-driven assessments that map your organizational maturity against industry benchmarks against all of these dimensions.
If you want to find out more about how that works, go to bsuper.ai.
And when you fill out the Get Started form, mention Maturity Maps.
Again, that's bsuper.ai.
Welcome back to the AI Daily Brief.
Today, we are connecting the dots between two stories.
The first is an announcement from the White House last week, invoking the Defense Production Act around U.S.
grid infrastructure.
And the second story is the release of the much-anticipated DeepSeek V4.
Still, to set up both of these stories, I want to go back a little bit.
to a discussion that really started to emerge in the last quarter of last year.
Taking you back to that time, you might remember that for much of Wall Street, the bubble narrative was reaching its peak.
GPT-5 had been disappointing, leading many people to trot back out the idea that seems to come up every six months or so in AI that we were hitting some pre-training wall.
And on top of that, we also had growing skepticism of the round of circular infrastructure deals that were driving and, according to some commentators, inflating future revenues.
Now looking back, it's become clear.
That the idea that we were going to have a bunch of excess compute sitting around, and that this massive buildout wasn't actually going to be necessary or justified by the market, seems to put it mildly a little quaint.
Instead, the narrative is shifting now to whether we're going to actually be able to provision as much compute as people want.
And upstream from all of that isn't just compute, it's the energy that that compute runs on.
Running counter to the tide at the time, Goldman Sachs called this out last year.
Summed up by Business Insider, they wrote that AI's next bottleneck wouldn't just be chips, but instead America's power grid.
GS identified that one of the biggest constraints to the future of AI development was going to be electricity itself.
Anticipating that data centers' share of electricity demand in the U.S.
would go from about 6% today to about double that to 11% by 2030, that shift, argued Goldman Sachs analysts, had the potential to cause untenable constraints on the U.S.
power grid.
As AI demands massive power, they wrote, a reliable and ample power supply is likely to be a key factor shaping this race, especially because power infrastructure bottlenecks can be slow to solve.
Now, the Financial Times, meanwhile, had written a similar story all the way back in December of 2025.
As they point out, boosting the U.S.
power grid is an enormous and time-consuming task due to a complex web of regulatory, financial, and supply chain challenges.
For example, they describe how the backlog of projects that are already waiting to plug into the grid have become, in their words, a major choke point.
And of course, there is the wider public dimension of this as well, where if we can't expand the total amount of power available, the hyperscalers instead have to try to consume as much as they can of what is currently available, which becomes a core issue with populations who face the potential of rising energy prices because of all of this.
Calls for efforts to get more aggressive about solving this problem have gotten louder as well.
At the end of March, J.P.
Morgan explicitly called our aging grid a national security risk and called upon the U.S.
government to do its part to solving what they are identifying as a major problem.
They write, It seems that the White House agrees.
At the beginning of the week, the White House posted a presidential memo for the Secretary of Energy.
It would be later in the week when the Internet started noticing and discussing the post.
In the memo, President Trump writes, Grid infrastructure and its associated upstream supply chains, including transformers, transmission lines and conductors, substantiations, high-voltage circuit breakers, power control electronics, protective relay systems, capacitor banks, electric core steel, and related raw materials and manufacturing tools, are industrial resources, materials, or critical technology items essential to the national defense.
2.
Without presidential action under Section 330 of the Act, United States industry cannot reasonably be expected to provide these capabilities for the needed industrial resource, material, or critical technology items in a timely manner due to limited domestic production capacity, extended procurement timelines, foreign supply dependence, and insufficient capital investment.
And three, purchases, purchase commitments, financial support for the development of production capabilities, or other action pursuant to Section 303 of the Act are the most cost-effective, expedient, and practical alternative methods for meeting this need.
I have declared a national emergency under Executive Order 14156, and I further determine that action to expand the domestic capability to develop, manufacture, and deploy grid infrastructure and supporting industrial supply chains is necessary to avert an industrial resource or critical technology item shortfall that would severely impair national defense capability.
You are authorized and directed to implement this determination, including making necessary purchases, commitments, and financial instruments to enable these projects.
TLDR, the White House is getting in the grid business, even if we don't know exactly what that means yet.
Most of the initial response from commentators had to do with the market.
TikTok Tick writes, grid is national security just in from the White House.
Expect a monster gap up in utilities next week.
Rosanna Prestia writes, companies working to electrify America will have a big tailwind.
And Citrini Research wrote a flash note exploring similar themes.
And of course, underlying all of this is perception of AI competition specifically with China.
Over the last 15 months, no company has done more to reshape the perception of that battle than DeepSeek did at the beginning of 2025.
When they released their free reasoning model, which rocketed to the top of the App Store charts and represented the first time that many had used a reasoning model, it wiped out an incredible amount of value from the U.S.
markets as beliefs about how far behind China was were totally reset.
Now, subsequent to that, over the course of 2025, Chinese open-weight models became a significant part of the AI landscape.
While they were never state-of-the-art, the fact that they were, frankly, nipping at the heels of state-of-the-art performance and doing so in a much cheaper package meant that many companies, especially startups with some amount of flexibility, were already finding integrations where they used state-of-the-art models from OpenAI or Anthropic or Google for hard and planning-type tasks, and models like DeepSeek or Quen or Kimi for a lot of the workhorse stuff that would cost a ton in resources otherwise.
Because DeepSeek was the first to really kick off that conversation, there has been extra hype around the release of their latest model, V4.
It's also been coming forever at this point.
I feel like a half dozen weeks in 2026 where people were sure it was coming this week, but at the end of last week, we did finally get it.
So in terms of the details, the V4 model family consists of two models.
There is a 1.6 trillion parameter model called V4 Pro, and a smaller version called V4 Flash.
Both versions have a million token context window, and on the benchmarks, V4 Pro is in a similar realm to Opus 4.5, 4.6, and GPT-5.3 and 5.4.
Sweebench Verified is basically a dead heat at around 80%.
On Terminal Bench 2.0, V4 Pro is slightly ahead of Opus 4.6 and slightly behind GPT-5.4.
And on Humanity's Last Exam, V4 Pro is slightly behind the Western comparisons.
And when it came to interpretations, there was frankly kind of a range.
Bloomberg's take was that the model was underwhelming.
They pointed to comments from Chris McGuire.
a senior fellow for China and Emerging Technologies at the Council on Foreign Relations, who wrote, It is not competitive with frontier U.S.
models and does not appear to close the gap with the United States in AI.
Former Trump AI advisor Dean Ball writes, R1 remains the closest I've seen Chinese models get to the U.S.
frontier.
That, by the way, is the one we were talking about from early 2025.
V4, continues Dean, is further behind than that, though that does not render it a useless or bad or uninteresting artifact.
In terms of the Twitterati, Leo Synthwave writes, My first impressions on DeepSeek V4, little disappointing that it's not state-of-the-art after all this time, but it's close.
New Pareto Frontier, a lot cheaper than 5.4 and Opus 4.6 for comparable performance.
Leo also did call it their new favorite model for creative writing.
Max Weinbach writes, Yeah, DeepSeek V4 Flash Pro don't really perform that well compared to any of the major US models, even 1-2 revisions old.
Looks like it's slightly behind Opus 4.5 in practice, and on par or slightly behind Kimi K2.6.
Some good optimization techniques there, but overall, eh.
Where the analysis gets a little more interesting is when it broadens out to incorporate price.
DeepSeq is pricing the Pro model at $174 per million input tokens and $348 per million output tokens.
That is less than a seventh of the cost of Opus 4.6 and less than a quarter of the cost of GPT-5.4.
V4 Flash is $0.14 per million inputs and $0.28 per million outputs, which undercuts Gemini Flashlight by 80%.
DeepSeq is even pricing Pro at around 25% lower than Kimi K2.6.
Simon Willison summed it up in his blog post as almost on the frontier, a fraction of the price.
Chinese AI analyst Po Zhao noted that DeepSeq said they are currently limited by compute supply and will drop prices even more once Huawei production is ramped up in the second half of the year.
He added, DeepSeq is publicly tying its API economics to domestic chip infrastructure.
That's the real headline.
One person who's taking all of this much more seriously than some of the initial dismissals is Matthew Berman.
He wrote a post called DeepSeek v4 is a serious threat.
And here's his main point.
Matt writes, here's the thing.
Most use cases don't require absolute frontier intelligence.
The vast majority of companies aren't doing frontier scientific research or trying to crack the hardest coding problems in the world.
They're running a business.
So imagine you're a CEO.
You'll look at GPT-55 at $30 per million output tokens or Opus 4.7 similarly priced.
Then you look at DeepSeq and it's a fraction of that.
It does almost everything you actually need.
It's open source so you can fine-tune it, host it how you like, control it precisely.
The calculus becomes really obvious.
Why would you pay so much more?
That's where the problem comes in.
Jensen Huang's been saying China is going to build their own chips and their own models, so they might as well be built on American technology, i.e.
NVIDIA chips.
Fine.
But the same argument now works in reverse.
If U.S.
enterprise companies build their AI strategy on top of Chinese open source models, that's a big geopolitical security risk.
If those Chinese AI labs changed their architecture or cut us off, we're suddenly in a really bad spot.
Matthew's argument is that, in his words, the U.S.
needs to go much harder on open source.
And second, even if we stay closed source, OpenAI and Anthropic need to get much cheaper much more quickly.
TLDR, DeepSeek didn't catch up to America, but they built something good enough, gave it away for free, and a lot of U.S.
companies are going to take them up on it.
Finally, almost as if to put a fine point on the geopolitics of these announcements, Around the same time, Beijing has launched into a flurry of activity to protect their national interests on AI.
On Friday, the same day DeepSeek v4 came out, Bloomberg reported that China plans to curb U.S.
investment in domestic tech companies.
The report stated that Chinese officials are putting out the word that Chinese tech firms should reject U.S.
capital unless explicitly approved.
The report named Moonshot, which is the developer of the Kimi models, as well as StepFun as two startups that had received these instructions.
Writes Bloomberg.
The overarching intent of the latest restrictions is to prevent U.S.
investors from taking stakes in sensitive sectors where national security is a priority.
The decision also follows a recent policy change that prevented foreign incorporated companies from going public in Hong Kong, cutting off a decades-old playbook for Chinese tech firms.
Instead, firms are shutting down their overseas corporations and reincorporating onshore.
Now, these changes were clearly a reaction to Meta's $2 billion acquisition of Manus, which has been under investigation in Beijing for several months.
Manus had moved their headquarters from Beijing to Singapore shortly before the deal was formed, which Beijing did not like.
In March, two Manus co-founders attended questioning in China and were told that they would not be permitted to leave the country while the investigation was ongoing.
On Monday, Beijing brought the investigation to its conclusion and blocked Meta's acquisition of Manus.
The Office of the Working Mechanism for Foreign Investment Security Review, an agency within the powerful National Development and Reform Commission, said in a statement, The office requires the parties involved to terminate and revoke the acquisition transaction.
The statement didn't give any specific reasons to unwind the deal, but cited national security grounds.
Chinese officials told the Financial Times that the deal was viewed as a conspiratorial effort to drain China of AI talent and resources.
Now, Meta is yet to comment on the decision, but this could get messy very quickly.
As many commentators have pointed out, a lot of these checks have already been cashed.
The Manus team already has Meta IDs.
How this all plays out in practice remains to be seen, but you take all these things together and we are clearly entering a new phase of the China-US AI competition.
I will certainly be watching to see if and what the Secretary of Energy actually does with this new power granted by the White House.
For now though, that's going to do it for today's AI Daily Brief.
Appreciate you listening or watching as always, and until next time, peace!