Gavin Baker on Invest Like The Best RE: Why the next massive data center might be… in space?

We kept seeing quotes from this interview w/ tech investor Gavin Baker on Patrick O'Shaughnessy’s Invest Like The Best pod being shared around X all week, so we thought we’d take a watch and see what all the fuss is about. And fuss about, we found.

Why you should care: Because Gavin cares A LOT about AI, and as an avid investor in the space, he knows A TON about every aspect of the industry from the chips to the interface layer.

The stand out section for us (besides the space data center stuff, though Jen Zhu has notes on that) argues that we are reaching a point of diminishing returns on raw “intelligence.” We break all of that down below, and highlight our favorite parts.

First, the TL;DR:

If you only have 3 minutes, read this: Tech investor Gavin Baker dropped some mind-bending predictions recently on the brilliant Invest Like The Best podcast from Patrick O'Shaughnessy, arguing that the laws of physics will inevitably push AI data centers into orbit—and that traditional software companies are walking into a death trap.

However, Baker's takes on AI ROI and the "Usefulness" gap is what really stood out to us. He argues that we are reaching a point of diminishing returns on raw "intelligence" (which we agree on). Unless you are asking deep questions about semiconductor physics, it is getting hard to tell the difference between the top models. So the next phase of AI isn't about IQ... it’s about Utility.

For example, he says the transition from "Smart Chatbot" to "Useful Agent" is what's needed now, and it relies on three specific building blocks:

• Massive Context: Usefulness requires memory. If you ask an AI to book a vacation, it shouldn't just know "beach." It needs to know you follow Andrew Huberman (so you need an East-facing balcony for morning sunlight) and that you refuse to fly on planes without Starlink.
• Reliability: The model can't just hallucinate a flight time. It needs to be boringly consistent.
• Task Length: We are moving from "Make me a reservation" (saves 5 mins) to "Plan this entire trip for my extended family" (saves 5 hours).

WHY IT MATTERS: This is the "ROI handoff." Raw intelligence is cool, but Context is the moat. Baker predicts that as context windows expand, AI will eventually hold every Slack message, email, and company manual you’ve ever written in its working memory... turning it into the ultimate Chief of Staff.

THE "SAAS DEATH TRAP": Baker also issued a grim warning for your favorite software stocks (think Salesforce or Adobe).

• Traditional software companies love their 80% profit margins.
• AI requires massive compute, meaning AI services usually have much lower margins (35-40%).
• The Trap: If big SaaS companies refuse to lower their margins to offer AI agents, lean AI startups will undercut them and steal their customers. Baker calls this a "life or death decision" that almost everyone except Microsoft is failing.

WHY IT MATTERS: We are currently in a hardware transition gap. NVIDIA’s new "Blackwell" chips were delayed, which should have stalled AI progress.

• The Savior: "Reasoning models" (like OpenAI's o1) saved the industry. They allowed AI to get smarter by "thinking longer" (Test-Time Compute) rather than just needing bigger chips.
• What’s Next: Elon Musk’s xAI is building data centers faster than anyone and will likely release the first model trained on these new super-chips in early 2026.

So why space datacenters? According to Baker, we are approaching a world where "scaling laws" (the rule that more chips = smarter AI) are colliding with Earth’s limitations. His solution? Space.

• Infinite Power: The sun is 30% more intense in space and shines 24/7. No batteries required.
• Free Cooling: Cooling is the most expensive part of a data center. In space, you just point a radiator at the dark void (which is near absolute zero) and heat vanishes for free.
• Faster Speeds: Lasers travel faster through the vacuum of space than they do through fiber optic cables on Earth.

WHAT TO DO: If you are a founder or executive at a software company, stop obsessing over high margins. Launch your AI agents now, even if they are expensive to run. If you don't, a startup will happily burn cash to steal your market share. For investors, keep an eye on xAI—they are currently doing the heavy lifting of debugging NVIDIA’s newest tech for the rest of the world at record pace and scale.

Now, lets dive into all that with more detail.

Top Takeaways from the Episode

AI Evaluation, Scaling Laws, and Hardware Transitions

• (1:48) The "Free Tier" Fallacy: Many investors judge AI capabilities based on free tiers (e.g., GPT-3.5), which is akin to judging an adult's capability by looking at a 10-year-old. To understand the frontier, you must pay the monthly fee for the highest tier (Gemini Ultra, Super Grok, etc.).
• (2:36) The "Vibes" of Research: AI development happens on X (Twitter). The "vibes" and public spats between research labs (like PyTorch at Meta vs. Jax at Google) are leading indicators of progress, often forcing lab leaders to intervene publicly.
• (5:14) Empirical Nature of Scaling Laws: Gemini 3 was crucial because it confirmed scaling laws for pre-training remain intact. However, we don't know why they work. We are like ancient Egyptians who could measure the sun's movement perfectly (aligning Pyramids) without understanding orbital mechanics.
• (6:28) The "Air Gap" in Progress: Without the emergence of "Reasoning" models, AI progress would have flatlined in 2024 and 2025. This is due to the hardware gap between the limits of Nvidia's Hopper architecture (max 200k coherent GPUs) and the delayed arrival of Blackwell.
• (7:33) Blackwell's Complexity: The transition from Hopper to Blackwell is the most complex product launch in tech history. It requires a shift to liquid cooling, racks increasing from 1,000 lbs to 3,000 lbs, and power requirements jumping from 30kW (30 homes) to 130kW.
• (9:17) The New Scaling Laws: Progress since late 2024 is driven by two new scaling laws: 1) Reinforcement Learning with Verified Rewards (RLVR) and 2) Test Time Compute. This allowed intelligence to jump from 8% to 95% on benchmarks like ARC AGI in just months.
• (10:23) Google's Temporary Advantage: Due to Blackwell delays, Google has a window of advantage with TPU v6 and v7. While Blackwell is an "F-35," early versions are hard to deploy, whereas Google is currently operating "F-4 Phantoms" (TPUs) but doing so efficiently at scale.

The Economics of Chips and Data Centers

• (11:18) Low-Cost Producer Paradigm Shift: For the first time in tech history, being the low-cost producer matters. Microsoft and Apple didn't win on low cost, but in AI, Google has been sucking the "economic oxygen" out of the room by producing tokens cheaper than anyone else.
• (12:11) xAI's Blackwell Lead: xAI (Elon Musk) will likely deploy the first Blackwell model in early 2026 because they build data centers faster than anyone else, acting as the "bug testers" for Nvidia's new architecture.
• (13:46) The GB300 Compatibility: The next Nvidia chip, GB300, is drop-in compatible with GB200 racks. This means companies mastering the complex GB200 liquid-cooled racks now will dominate later as the low-cost producers when they slot in the upgraded chips.
• (14:39) Google's Strategic Crisis: Once Blackwell scales, Google may lose its low-cost advantage. This forces a difficult decision: continue running AI at negative 30% margins to starve competitors, or raise prices and lose share.
• (16:32) The Broadcom Threat: Google pays Broadcom estimated 50-55% gross margins for TPU backend design. As spending hits ~$30B, it becomes economically inevitable for Google to bring this in-house to save billions, threatening Broadcom's revenue.
• (19:21) Nvidia's "One Year" Defense: Nvidia's defense against hyperscalers (Amazon/Google) building their own ASICs is accelerating their roadmap to a one-year release cadence. It is incredibly difficult for internal ASIC teams to keep up with that velocity.
• (20:00) The ASIC Learning Curve: It takes three generations to build a competitive chip. Amazon's ASIC team is the best among hyperscalers (Graviton/Trainium), but generally, internal chips lag behind Nvidia's rapid iteration.

ROI, Utility, and Bear Cases

• (22:30) Utility over Intelligence: The next phase of AI isn't just "smarter" models, but cheaper tokens allowing models to "think" longer. This enables agents to hold massive context (e.g., family preferences, logistics) to execute complex tasks like booking a full vacation, not just answering a query.
• (24:01) Verification is Automation: Any task that has a verifiable outcome (Did the code compile? Did the ledger balance? Did the sale close?) will be mastered by AI using reinforcement learning.
• (26:03) Positive ROI is Visible: The "ROI question" is settled for public companies. Major spenders on GPUs report higher ROIC and revenue per employee now than before they ramped spending.
• (27:40) The Edge AI Bear Case: The most plausible bear case for the massive GPU build-out is "Edge AI." If a pruned model on an iPhone (115 IQ) running free is "good enough," it could destroy the economics of massive cloud models.
• (28:50) Context as the Moat: The defense against Edge AI is Context Windows. If the cloud model holds every email, Slack message, and document a company has ever produced, the "Edge" model cannot compete regardless of its IQ.

Market Dynamics & Geopolitics

• (40:38) Reasoning Creates a Flywheel: Pre-training had no data flywheel (deploying a model didn't automatically make the next one better). "Reasoning" models change this: verifiable answers create data that feeds back into training, creating increasing returns to scale for the top labs.
• (42:58) Big Tech Failures: Despite nearly unlimited resources, Meta, Microsoft, and Amazon failed to build a top-tier frontier model in 2024. This proves that building frontier AI is much harder than simply throwing capital at the problem.
• (44:17) GPU Utilization Variance: There is a massive, under-discussed gap in competence. Some companies run GPU clusters at 30% uptime, while leaders run them at 90%. This operational "skill issue" is a major differentiator.
• (46:52) Chinese Open Source: Chinese open source models are a "gift from God" to Meta, allowing them to bootstrap their own models using Chinese checkpoints.
• (47:33) The Rare Earth Miscalculation: China's restriction on rare earths was a strategic mistake. It forced the US to innovate alternative refining methods (biological/enzymatic) and supply chains, which will solve the shortage faster than anticipated.
• (48:27) The Compute Gap: DeepSeek's technical papers subtly admit they cannot compete with US labs due to a lack of compute (the Blackwell ban). As Blackwell rolls out, the gap between US frontier models and Chinese models will blow out significantly.

The SaaS Mistake and Investing Lessons

• (1:11:32) The SaaS Death Trap: Application SaaS companies (Salesforce, Adobe) are making the same mistake brick-and-mortar retailers made with e-commerce. They are addicted to 80-90% gross margins and refuse to embrace AI Agents because AI runs at ~40% margins.
• (1:12:37) The Activist Playbook: There is a massive opportunity for activist investors to force SaaS companies to accept lower margins to survive. If they don't, AI-native startups willing to run at 35% margins will wipe them out.
• (1:20:01) Investing as Truth Seeking: Investing is the search for hidden truths. You generate alpha by finding a truth before others see it.
• (1:22:30) The "Ski Bum" Origin: Baker originally planned to be a ski bum, river guide, and wildlife photographer. His parents forced him to take one internship, which led him to read Peter Lynch and Warren Buffett, instantly changing his life trajectory.
• (1:24:46) History + Current Events: Investing is the intersection of a thorough knowledge of history and an accurate understanding of current events to predict the next move in a game of skill and chance.

The Future: Space Data Centers and Energy

• (52:17) Prediction: Data Centers in Space: The most important development in the next 3-4 years will be orbital data centers. From a first-principles physics perspective, space is the optimal location for compute.
• (52:47) Three Advantages of Space: 1) Power: Solar is 30% more intense and available 24/7 (no batteries needed). 2) Cooling: Radiating heat into deep space (near absolute zero) is "free," removing the massive HVAC cost of Earth data centers. 3) Speed: Laser communication in a vacuum is faster than light through fiber optic cables.
• (55:51) Direct-to-Cell Latency: With Starlink direct-to-cell, a query goes from phone -> satellite -> phone. This bypasses the cell tower -> fiber -> metro aggregation -> datacenter route, offering lower latency and better user experience.
• (56:41) The Musk Convergence: Tesla (Optimus bodies/Vision), xAI (Intelligence/Brains), and SpaceX (Launch/Orbital Compute) are converging into a single integrated advantage where each strengthens the others.
• (58:31) Monetization via Ads: Google has begun monetizing "AI Overviews" with ads. This gives permission for the entire industry to introduce ads into free-tier models, creating a massive new revenue stream for OpenAI and others.
• (1:00:31) TSMC's Caution: TSMC is acting as a "governor" on the bubble by refusing to expand capacity too fast (fearing a crash). This paradoxically extends the cycle and keeps pricing power high for chipmakers.
• (1:02:02) Power Constraints Negate Price: When power (watts) is the hard constraint, the price of the chip is irrelevant. Buyers will pay anything for the chip that delivers the most tokens per watt.‍
• (1:03:03) Nuclear is Too Slow: We cannot build nuclear power fast enough in the US due to regulations (NEPA). The actual solution for AI power demand is Natural Gas (fracking) and Solar.

Now let's break down our favorite parts in more detail...

The Pivot: Why AI is moving from "IQ" to "Utility"

We have spent the last two years obsessing over benchmarks. Does this model score 85% on the Bar Exam or 90%? Can it solve a physics problem better than a grad student?

According to Gavin Baker, for the average user, this "IQ race" is starting to hit a wall of indifference.

If you pay for the top-tier models (Gemini Ultra, Grok Super, GPT-4o), they are all essentially "geniuses." Unless you are an expert in a niche field—Baker uses the example of asking detailed questions about "PCI Express vs. Ethernet protocols"—you probably can't tell the difference between them anymore.

The "Usefulness" S-Curve‍

Baker argues we are at a critical handoff point. We need to stop optimizing solely for Intelligence and start optimizing for Usefulness.

Intelligence is knowing the answer. Usefulness is doing something with it.

The Building Blocks of Usefulness

So, how do we get there? Baker identifies the specific components that turn a chatbot into a useful agent:

1. Extreme Context (The "Huberman" Rule)‍

This is the biggest technical bottleneck. To be useful, an AI needs to know you, not just general knowledge.
Baker gave a specific personal example: If an AI is going to be his travel agent, it isn't enough to just book a hotel. It needs to hold a massive amount of personal context in its "brain" simultaneously:

• The Huberman Preference: "I follow Andrew Huberman, so I need an East-facing balcony to get my morning sunlight."
• The Connectivity Preference: "I need to be on a plane that has Starlink."
• The Family Web: It needs to know the preferences of his parents, his sister, and his niece and nephew, and how those preferences conflict.

2. Task Duration (The "Thinking" Time)

‍Currently, AI is good at short bursts of work (e.g., "Write this email"). Usefulness comes from extending that duration.

• Level 1: "Book me a table at 7 PM." (Economically useful, but low value).
• Level 2: "Plan my entire vacation, coordinate the flights, and ensure the hotel rooms match my weird sunlight preferences." (High value, saves a human 3–4 hours of work).

3. Verification & Reliability‍

This is the final block. You can't have an agent booking flights if it hallucinates the departure time. Baker notes that AI will first conquer industries where the outcome is verifiable (like accounting or coding) before it masters subjective tasks.

The "Air Gap" and How Reasoning Saved AI

A major revelation from Baker is that, technically, AI progress should have stalled in late 2024.

The industry is currently undergoing the most complex hardware transition in history: moving from NVIDIA’s "Hopper" chips to the new "Blackwell" architecture. This isn't just a chip swap; it requires a complete overhaul of data center physics. Racks are going from weighing 1,000 lbs to 3,000 lbs, and power requirements are jumping from 30 kilowatts to 130 kilowatts per rack.

Because of this complexity, Blackwell was delayed. Under normal circumstances, this hardware plateau would have frozen AI intelligence levels for 18 months.

So why are models getting smarter?‍

Baker argues that "Reasoning Models" (like OpenAI’s o1) saved the industry. By introducing "Test-Time Compute"—allowing a model to "think" for seconds or minutes before answering—labs were able to increase intelligence without better hardware.

"Reasoning bridged the gap," Baker explains. It allowed the industry to survive until the heavy artillery (Blackwell) arrives in volume in 2026.

The "Egyptian" Theory of Scaling Laws

When Google’s Gemini 3 was released, Baker says it proved one critical fact: "Scaling laws for pre-training are intact."

This is a big deal because, as Baker points out, "no one on Planet Earth knows how or why scaling laws for pre-training work." He used a fantastic analogy to explain our current understanding of AI:

"Our understanding of scaling laws… is kind of like the ancient Egyptians' understanding of the sun. They could measure it so precisely that the Great Pyramids are perfectly aligned… but they didn't understand orbital mechanics. They had no idea 

The "New" Laws:‍

Baker argues that we actually have three scaling laws stacking on top of each other now.

1. The Old Law (Pre-Training): More chips + more data = smarter model. This is still working (proven by Gemini 3).
2. The New Law (Post-Training): "Reinforcement Learning with Verified Rewards." This is the reasoning layer (like OpenAI's o1).
3. Test-Time Compute: Allowing the model to "think" longer to get a better answer.

Baker revealed that without those new reasoning laws, AI progress would have flatlined in late 2024. Why? Because we hit a physical limit on the old hardware.

• The Limit: You currently cannot get more than ~200,000 Nvidia Hopper GPUs "coherent" (working as a single brain) before the physics break down.

The Save: Because the next-gen chips (Blackwell) were delayed, "Reasoning kind of saved AI." It allowed the models to get smarter by thinking longer, rather than just being bigger.

Now, for the first year, in 2026, these chips will be used for Training (making models), not Inference (running products).

FYI, from a business POV, you can look at the difference between training and inference like so:

• Training = All cost, no revenue.
• Inference = Where the money is made.

The "Receipts": Proof that AI is actually making money (ROI)

There is a loud narrative right now that AI is a money pit—billions in CapEx with zero return. Baker says that is empirically false, and he brought the audited financials to prove it.

According to Baker, the largest buyers of GPUs are public companies. Because they are public, we can calculate their Return on Invested Capital (ROIC). The result? The ROIC of the big public spenders is higher today than it was before they started ramping up their AI spending.

How AI fixed a trucking giant (and saved the market).

Baker’s answer to the haters? The returns are real, and they are showing up in the most boring places imaginable.

Baker pointed to C.H. Robinson, a Fortune 500 logistics company, as the ultimate "mic drop" example for anyone questioning the ROI of AI.

The C.H. Robinson Case Study:

C.H. Robinson matches shippers (people who need to move stuff) with truckers (people with empty trucks). A key part of their business is quoting prices to customers.

• Before AI: It took human agents 15 to 45 minutes to generate a price quote. Because of this slowness, they only responded to about 60% of inbound requests.
• After AI: They are now quoting 100% of inbound requests, and they are doing it in seconds.
• The Result: The company posted such a strong quarter driven by these productivity gains that their stock jumped ~20% on the earnings news.

Why this matters for the Fortune 500

‍Baker argues this is a massive signal. Historically, Fortune 500 companies are dinosaurs—they are the last to adopt new tech.

• The Cloud Era: When the Cloud launched, startups adopted it instantly. It took the Fortune 500 nearly 5 years to standardize on it.
• The AI Era: This time, it’s different. We are already seeing "boring" non-tech companies (freight, banking, industrial) reporting hard numbers on AI productivity.

So, to recap, Baker was worried that the massive spending gap between scaling and training new models crash the markets. But because companies like C.H. Robinson are finding efficiency gains now (using current tech), they are proving that AI can pay for itself immediately. As Baker puts it, if an old-school trucking company can figure this out, the "ROI Air Gap" might not be so scary after all.

The "Verification" Rule: If you can grade it, AI can master it.

Gavin Baker dropped a golden rule for understanding what AI will automate next, borrowing a concept from Andrej Karpathy: "With software, anything you can specify, you can automate. With AI, anything you can verify, you can automate."

The logic is simple: If there is a clear "Right" or "Wrong" outcome, you can train an AI using Reinforcement Learning (trial and error) to master it perfectly. If the outcome is subjective (like "write a funny poem"), it’s much harder.

Here are the specific examples Baker gave of jobs that are about to get solved because they are easily verified:

• Accounting & Finance: Does the model balance? Do the global books reconcile? In double-entry bookkeeping, the math either works or it doesn’t. Since the outcome is binary, AI will likely master accounting before it masters creative writing.
• Sales: Did the deal close? Yes or No. Because the "win" condition is undeniable, AI agents can simulate millions of sales calls to figure out the perfect persuasion tactics.
• Customer Support: Did the customer ask for an escalation to a human manager? If yes = Fail. If no = Pass.
• Gaming (AlphaGo): Did you win the match? This is why AI crushed humans at Chess and Go first—the scorecard is absolute.

If your job involves tasks where there is an undeniable "correct" answer at the end of the day, you are first in line for automation.

The "SaaS Math" Warning:‍

While ROI is positive for companies that use AI, Baker warns that companies selling AI software (SaaS) are facing a mathematical crisis regarding their margins.

• The Old Math: Traditional SaaS companies (like Adobe or Salesforce) are used to 70–90% gross margins. They write code once and sell it cheaply.
• The New Math: AI Agents require constant compute to "think." A healthy AI-native company typically runs at ~40% gross margins.
• The Warning: Baker argues that if legacy software companies try to preserve their 80% margins, they are "guaranteeing" failure. They must be willing to accept lower margins to compete, or AI startups (who are happy with 40% margins) will undercut them.

The AI Bear Case: What if the iPhone is "Good Enough"?

While Baker is undeniably bullish, he outlined what he calls the "most plausible and scariest bear case" for the AI boom. It isn't regulation or safety—it’s "Edge AI."

The Scenario:‍

What happens if, in 3 years, Apple or Samsung can run a compressed, "pruned down" version of a top-tier model directly on your phone?

Baker paints a specific picture of this future device:

• It runs a model equivalent to Gemini 5 or Grok 4 (a "God model").
• It runs at 30 to 60 tokens per second (fast reading speed).
• It has an IQ of ~115 (smart enough for most tasks).

The Threat:‍

If that exists on your phone, "Then that’s free."
Baker argues this is clearly Apple’s strategy: Be the distributor. Your iPhone handles 95% of your questions for free using a "good enough" local model, and only calls out to the massive, expensive "God models" in the cloud for the really hard stuff.

If people are satisfied with the free, 115-IQ model in their pocket, the demand for massive $100B server farms in the cloud could crash.

The Counter-Argument (Why the Bear Case might fail):‍

Baker says the only defense against this "Edge AI" takeover is Context.

• To be truly useful, an AI needs to know everything about you—every Slack message, email, PDF, and preference (e.g., "I like morning sun in my hotel room").
• Phones have limited memory. If cloud models can hold millions of tokens of context (your entire digital life) in their working memory, they will always be smarter and more useful than the "amnesiac" model on your phone.

The Geopolitics of AI: Why China’s "Rare Earth" gambit is about to backfire.

Baker didn’t mince words when it came to the AI cold war. His take? China has made a massive strategic miscalculation, and the "gap" between American and Chinese AI is about to turn into a chasm.

The "Gift" and the "Trap"‍

Baker started by acknowledging that Chinese open-source models (like DeepSeek) are currently very impressive. In fact, he called Chinese open source "a gift from God to Meta," because companies like Meta can use those models to bootstrap their own research.

But that dynamic is about to change fast.

The "Blackwell" Wedge‍

The arrival of Nvidia’s new Blackwell chip is the geopolitical game-changer.

• The US Advantage: American labs (xAI, OpenAI, Google) are getting Blackwells—which are exponentially faster and more efficient.
• The Chinese Disadvantage: China is stuck using domestic chips (like Huawei’s Ascend) or older, smuggled Nvidia chips. Baker compares this to fighting a war with F-4 Phantoms while the US has F-35s.

Baker noted that DeepSeek (a leading Chinese AI lab) actually admitted in a recent technical paper that they are struggling to compete because they simply "don't have enough compute."

The "Rare Earth" Miscalculation‍

China recently restricted exports of Rare Earth minerals (essential for making chips), thinking this gave them leverage over the US. Baker calls this a "terrible mistake."

• The Reality: Technology > Dirt. Baker argues that China will soon realize, "Whoopsie daisy. We need the Blackwells."
• The Fix: Baker predicts the US will solve the rare earth shortage faster than people think (thanks to new DARPA tech and refining methods in friendly countries).

China thought they could starve the US of minerals. Instead, the US is starving China of intelligence. As Blackwell rolls out in 2026, Baker predicts the performance gap between American AI and Chinese AI is going to "blow out," leaving China further behind than they realized.

The Final Frontier: Data Centers in Space

Baker’s most "sci-fi" prediction is actually based on cold, hard economics. He argues that in the next 3-4 years, we will see a gold rush for data centers in orbit.

Why? First Principles:

• Power: Solar energy is 30% more intense in space (no atmosphere) and available 24/7 (no night). You don't need massive battery backups.
• Cooling: On Earth, cooling is the most expensive and difficult part of running a GPU cluster. In space, you simply point a radiator at the dark side of the void (which is near absolute zero), and cooling is free.
• Speed: Light travels faster in a vacuum than it does through glass (fiber optic cables). A mesh network of laser-linked satellites could technically process and transmit data faster than a terrestrial network.

With SpaceX’s Starship lowering launch costs, the economics of space-based compute are starting to pencil out.

The Chip Wars: xAI vs. Google vs. Nvidia

As far as who is positioned to win the AI wars, Baker says we are witnessing a brutal game of 4D chess between the tech giants.

• Google’s Strategy: Google is currently the "low-cost producer" of AI tokens (using their custom TPU chips). Baker suggests Google is weaponizing this by keeping prices artificially low to "suck the economic oxygen" out of the room, starving competitors who rely on VC funding.
• The xAI Factor: Elon Musk’s xAI is building data centers faster than anyone else. Baker predicts xAI will be the first lab to release a model trained on NVIDIA’s Blackwell chips, likely in early 2026.
• The Equalizer: Once the next version of Blackwell (the GB300) hits the market, the cost to produce tokens will plummet for everyone, neutralizing Google’s advantage and kicking off the next leg of the race.

With 2026 right around the corner, we'll see if what he predicts will come to pass soon.

BTW, if you haven't watched Patrick's interview with Dylan Patel, probably the only other encyclopedically obsessed and knowledgable expert covering the entire AI industry, you need to watch that next.