How Deepseek v4 Connects to the US Power Grid

April 27, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
April 27, 2026

Sign up for AgentOS, our latest free training course. It’s a platform and model agnostic program to help you build a personal agentic operating system that can evolve with you over time. The program is self-directed and project based, running over four weeks.

HEADLINES

Google to Invest Up to $40B in Anthropic

Google has expanded its Anthropic relationship with another massive investment — $10B upfront and another $30B contingent on undisclosed commercial milestones, on top of the 5GW Broadcom-built TPU deal from last month. The structure looks a lot like the Amazon deal from earlier in the month: equity for compute. Citrini analyst Jukan flagged a Mirae Securities note as one of the most interesting takes, arguing that all of Anthropic's unusual moves around the GPT-5.5 launch converge on a single conclusion — to secure compute, Anthropic must bind itself far more deeply to those who possess the physical resources. Each gigawatt is roughly a full-scale nuclear reactor, and Microsoft's entire 2024 global data center footprint was around 6GW; Anthropic alone is now locking in incremental capacity that rivals the entirety of Microsoft's historical physical infrastructure. Some read this as Google giving up on Gemini, others as a hedge or an entirely new business line — and on the deal terms (anchored to the $350B February valuation rather than the $800B secondaries), Google is extracting a healthy premium for that compute.

Meta Signs Multi-Billion Dollar Deal for Amazon's Graviton CPUs

Meta has signed a multi-billion dollar deal to rent AI chips from Amazon — but with an interesting twist. Rather than signing up for Trainium ASICs, Meta is renting Amazon's Graviton 5 CPUs, which are optimized for agentic workloads. There's a growing conversation around whether CPU architecture could prove more efficient than GPUs when it comes to actually running agents. Given Meta's entire strategy is centered on delivering consumer agents, the commitment to agent-specific architecture makes a lot of sense — though this could also just be Meta placing as many bets as possible and snapping up every chip it can get its hands on. Meta's AI buildout is forecast at up to $135B this year and they already rent from AWS, NVIDIA, AMD, Google, CoreWeave, and Nebius.

The AI Trade Is Back, NVIDIA Hits $5 Trillion

The S&P 500 is up 12.5% over the past month, completely erasing the Iran War drawdown — and the Wall Street Journal frames the recovery as all about AI. Of the 82 stocks up more than 10%, almost all are AI-related; without the Mag 7 plus Broadcom, the S&P 500 is actually down year to date. Even dotcom darling Cisco hit a new all-time high this week on the back of data center spending, taking 26 years to claw back to its 2000 peak. While the WSJ narrative still focuses on questioning the bubble, analysts are insisting this time is different — Benchmark's Cody Acree said the demand, the spending, and the CapEx budgets are real. NVIDIA closed the week as the world's first $5 trillion company, just nine months after becoming the first to $4 trillion. As one quote in the WSJ piece puts it: not enough people are emotionally prepared for if it's not a bubble.

MAIN STORY

How DeepSeek V4 Connects to the US Grid

Today we're connecting the dots between two stories: the White House invoking the Defense Production Act around US grid infrastructure, and the long-awaited release of DeepSeek V4. The throughline is China-US AI competition. The narrative that we'd have excess compute sitting around now looks quaint — instead the question is whether we can provision as much compute as people want, and upstream of compute is the energy that runs it. Take all this together and we are clearly entering a new phase of US-China AI competition.

AI NATIONAL SECURITY RISK GROWS

The Grid as the Real Bottleneck
"AI's next bottleneck won't be chips — it's the power grid."
Goldman Sachs called this out last year, projecting data centers will go from ~6% to ~11% of US electricity demand by 2030 — a shift that could cause untenable constraints. The FT made the same point back in December 2025: boosting the grid is enormous and time-consuming due to regulatory, financial, and supply chain challenges, and the backlog of projects waiting to plug in has become a major choke point. There's a public dimension too — if total power can't expand, hyperscalers consume what's available and consumers face rising energy prices.

JP Morgan
"Aging grids are a national security risk."
At the end of March, JP Morgan explicitly called the aging US grid a national security risk and called on the government to act. Electric grids are undergoing a fundamental reframe from aging legacy assets to strategic infrastructure that must withstand physical threats, technological change, and growing supply needs — grid resilience now underpins economic development, industrial competitiveness, and national security.

The White House Memo
"Grid infrastructure is now essential to national defense."
At the beginning of last week, the White House posted a presidential memo invoking Section 303 of the Defense Production Act — declaring that transformers, transmission lines, conductors, substations, high-voltage circuit breakers, power electronics, capacitor banks, electric core steel and related raw materials are "industrial resources, materials, or critical technology items essential to the national defense." The memo states that without presidential action, US industry cannot reasonably be expected to meet these needs in time due to limited domestic production capacity, extended procurement timelines, foreign supply dependence, and insufficient capital investment. TLDR: the White House is getting in the grid business, even if we don't know exactly what that means yet.

Market Reaction
"Expect a monster gap up in utilities next week."
Most of the initial commentary focused on the market angle — companies working to electrify America have a big tailwind, and Citrini's flash note explored the same themes. Underlying all of it is the perception of AI competition with China.

DeepSeek V4 Arrives
"It's been coming forever — finally, V4."
No company has done more to reshape perception of the US-China AI battle than DeepSeek did at the start of 2025 with R1, which wiped out a stunning amount of value from US markets overnight. Over the course of 2025, Chinese open-weight models became a significant part of the landscape — never state-of-the-art, but nipping at the heels and frankly cheap enough that startups were already routing workhorse tasks to DeepSeek, Qwen, or Kimi. V4 was a long time coming — feels like half a dozen weeks in 2026 where people were sure it was dropping that week — and at the end of last week we finally got it. The family is V4-Pro (1.6T params) and V4-Flash, both with 1M token context. On benchmarks, V4-Pro is in similar territory to Opus 4.5/4.6 and GPT-5.3/5.4 — basically a dead heat on SWE-Verified at ~80%, slightly ahead of Opus 4.6 and slightly behind GPT-5.4 on Terminal Bench 2.0, slightly behind on Humanity's Last Exam.

The "Underwhelming" Take
"V4 doesn't close the gap with US labs."
Bloomberg's take was that the model was underwhelming and failed to close the lead with US labs, pointing to comments from Chris McGuire (Senior Fellow for China and Emerging Technologies at the Council on Foreign Relations) that V4 is not competitive with frontier US models and does not appear to close the gap. Former Trump AI advisor Dean Ball wrote that R1 remains the closest he's seen Chinese models get to the US frontier — V4 is further behind than that, though that doesn't render it useless or uninteresting. Twitterati echoed it: synthwavedd called V4 "a little disappointing" but a new Pareto frontier and his new favorite for creative writing; Max Weinbach said it doesn't perform that well versus US models even one to two revisions old.

The Price Story
"Almost on the frontier, a fraction of the price."
Where the analysis gets more interesting is price. V4-Pro is $1.74/M input and $3.48/M output — less than a seventh the cost of Opus 4.6 and less than a quarter of GPT-5.4. V4-Flash undercuts Gemini Flash-Lite by 80%, and Pro is roughly 25% cheaper than Kimi K2.6. Simon Willison summed it up as "almost on the frontier, a fraction of the price." Chinese AI analyst Poe Zhao noted DeepSeek said they're currently limited by compute supply and will drop prices further once Huawei production ramps in the second half of the year — and added that DeepSeek is publicly tying its API economics to domestic chip infrastructure, which is the real headline.

Matthew Berman: V4 Is a Serious Threat
"DeepSeek didn't catch up. It built something good enough."
One person taking V4 much more seriously than the dismissals is Matthew Berman, who wrote that most use cases don't require absolute frontier intelligence — the vast majority of companies aren't doing frontier scientific research, they're running a business. Imagine you're a CEO looking at GPT-5.5 at $30/M output tokens or Opus 4.7 at similar prices, then at DeepSeek for a fraction of that, doing almost everything you actually need, open source so you can fine-tune and host it yourself. The calculus becomes obvious. Jensen has been saying China will build its own chips and models so they may as well be built on American technology — but the same argument now works in reverse. If US enterprise builds AI strategy on Chinese open source, that's a big geopolitical security risk. Berman's argument: the US needs to go much harder on open source, and even if labs stay closed, OpenAI and Anthropic need to get much cheaper much faster. DeepSeek didn't catch up to America — they built something good enough, gave it away for free, and a lot of US companies are going to take them up on it.

Beijing's Counter-Move: Manus Blocked, US Investment Curbed
"China just unwound a $2B Meta deal that's already done."
Almost as if to put a fine point on the geopolitics, Beijing has launched into a flurry of activity to protect their national interests on AI. On Friday — the same day V4 came out — Bloomberg reported China plans to curb US investment in domestic tech firms, with Moonshot (Kimi) and Stepfun named as startups already told to reject US capital unless explicitly approved. The overarching intent is to prevent US investors from taking stakes in sensitive sectors. The decision follows a recent change preventing foreign-incorporated firms from going public in Hong Kong — cutting off a decades-old playbook for Chinese tech firms, which are now shutting down overseas corporations and reincorporating onshore. These changes were clearly a reaction to Meta's $2B acquisition of Manus, which had moved its HQ from Beijing to Singapore shortly before the deal — something Beijing did not like. In March, two Manus co-founders were detained for questioning. On Monday, Beijing brought the investigation to its conclusion and blocked the acquisition outright, with the Office of the Working Mechanism for Foreign Investment Security Review requiring the parties to terminate and revoke the transaction. Officials told the FT they viewed the deal as a conspiratorial effort to drain China of AI talent and resources. Meta hasn't commented — but as many have pointed out, a lot of these checks have already been cashed. The Manus team already has Meta IDs. How this plays out in practice remains to be seen.