OpenAI Stalls, Tech Giants Burn $725B, Retatrutide Weight-Loss Drug Data Explodes: All In Podcast Deep Dive

When OpenAI's $600 billion compute bill nearly equals its valuation, you have to ask: are they building the future, or gambling with future money? In the May 2026 episode of the All In Podcast, four top Silicon Valley investors delivered a remarkably information-dense discussion covering OpenAI's stalling growth, tech giants' astronomical capital expenditures, AI cybersecurity, the Musk lawsuit, and the next-generation weight-loss drug Retatrutide. This article breaks it all down.

---

OpenAI Misses Both User and Revenue Targets; IPO Probability Cut in Half

The Wall Street Journal dropped a bombshell: OpenAI's target of reaching 1 billion weekly active users by end of 2025 still hasn't been met as of April 2026. The revenue side is equally disappointing—ChatGPT's 2025 revenue target was also missed, with current annualized revenue sitting somewhere between $20 and $30 billion.

That number sounds impressive, but compare it to OpenAI's $600 billion compute spending commitment—roughly equal to its entire secondary market valuation—and the severity of the problem becomes clear.

Even more telling are the internal divisions. CFO Sarah Friar has publicly expressed concerns about "spending outpacing revenue," clearly not on the same page as Altman regarding IPO timing. When a company's CFO starts speaking publicly like this, it's not caution—it's a distress signal to the market. On Polymarket, the probability of OpenAI going public before end of 2026 has plummeted from 60% to 32%. The market has voted with real money.

That said, Sacks offered a noteworthy contrarian view: on the product side, GPT 5.5 is performing incredibly well. Built on the entirely new foundation model Spud, it represents OpenAI's first true base model upgrade in over a year, and developer community feedback has been extremely positive. Codex is rapidly capturing market share in the coding space.

Meanwhile, competitor Anthropic isn't faring much better. Opus 4.7 has been caught throttling compute and compressing thinking time, with many users reverting to version 4.6. This exposes a structural contradiction across the industry: the stronger the model, the higher the inference costs, and paradoxically, the less companies dare to let users fully utilize it. The 1 billion WAU target was itself an absurd anchor—globally, products with over 1 billion daily active users can be counted on one hand, yet OpenAI treated it as a given milestone.

The core question remains: can developer enthusiasm translate into enough commercial revenue to cover a $600 billion compute commitment?

---

AI's Real Bottleneck Isn't GPUs—It's Electricity

Chamath hit the nail on the head with a fact almost everyone overlooks: AI's real bottleneck isn't compute—it's power. The entire chain runs electricity → compute → tokens → services, and electricity, the most fundamental layer, is becoming the chokepoint.

The data is staggering: fewer than half of announced data center projects are actually under construction, with most stuck in permitting and supply chain bottlenecks. Worse still, approximately 40% of announced AI projects may be canceled outright—partly because community opposition to AI and data centers is spreading.

You can exponentially scale algorithms, but you cannot exponentially scale the electrical grid. This is the physical reality that digital utopians least want to confront.

Chamath believes both OpenAI and Anthropic are trillion-dollar companies, but that hyperscale cloud providers (Oracle, Amazon, Meta, Microsoft, Google) are the true beneficiaries. He also specifically mentioned that Grok and SpaceX have significant idle compute capacity, suggesting Elon should collaborate with Dario Amodei.

However, there's another path. A recent MIT paper demonstrated that pruning techniques can shrink neural networks by 90%, directly reducing inference costs by 10x. Small Language Models (SLMs) and Vertical Small Language Models (VSLMs) are emerging as an entirely different competitive trajectory. If the efficiency revolution outpaces the compute arms race, today's giants madly spending on data centers may be constructing the Maginot Line of the 21st century.

Freeberg cited BCG's classic "Rule of Three"—mature markets tend to converge toward a 4:2:1 market share ratio. But before consolidation arrives, whoever burns through cash first exits first.

---

AI Cybersecurity: Elite Hacker Capabilities Are Being Democratized

OpenAI released GPT 5.5 Cyber, which after testing by the AI Safety Institute, became the second model capable of completing end-to-end multi-step cyberattack simulations—and it's commercially ready with no compute restrictions.

This was the most unsettling segment of the entire episode.

Sacks emphasized that neither Mythos nor GPT 5.5 Cyber creates vulnerabilities—they discover them—and should be viewed as defensive tools. The logic is sound, but reality is more complicated. Of the roughly 5 million cybersecurity professionals globally, only a few thousand are truly elite hackers. AI is mass-replicating those few thousand people's capabilities—and whether white hats or black hats use it first has never been a technology question, but an incentive structure question.

China's frontier models (such as DeepSeek 4) currently sit at approximately 80-85% of U.S. frontier capability, and are expected to reach parity within 6 months. The attack-defense balance in cybersecurity is about to be fundamentally shattered, and we haven't even established basic international rules frameworks.

Chamath also dropped an ominous hint: Palo Alto Networks (led by Nikesh Arora) already possesses the ability to penetrate and manipulate all AI models. If true, we're building our entire digital civilization on infrastructure that can be controlled by a handful of people.

Over the next 5 to 6 years, global operational software will be massively rewritten. Machine-written code may be more secure, but the nature of cyber threats will transform entirely into machine-versus-machine warfare.

---

Musk v. OpenAI: A Diary That Became a Legal Bomb

Elon is alleging that OpenAI violated its charitable trust and engaged in unjust enrichment by converting a nonprofit into a for-profit entity, seeking $150 billion in damages. His exact words: "If looting charitable organizations is allowed, the entire foundation of charitable giving in America will be destroyed."

The most dramatic twist comes from Greg Brockman's diary. Yes—OpenAI's co-founder kept a diary detailing exactly how they pushed Elon out and drove the conversion to for-profit. The diary states in black and white: "The real answer is we wanted Elon out" and "We were dishonest with him."

Freeberg's reaction on the show spoke for everyone—he directly quoted the classic line from HBO's The Wire: if you're going to do something wrong, never leave a paper trail.

Polymarket gives Elon roughly a 42-43% chance of winning. The most likely outcome is a technical victory for Elon where he only recovers his original $40 million donation—a moral victory for Elon, a legal victory for OpenAI, and a warning for the entire industry: your founding story will eventually become your legal liability.

Worth noting: OpenAI once offered Elon equity, but he refused because he insisted the entity should remain charitable. The presiding judge, Rogers, is a 61-year-old Obama appointee who previously oversaw Epic Games v. Apple. This is a bench trial; the jury serves only in an advisory capacity.

Sacks refused to comment on the show—last time he commented on an Elon lawsuit, he got subpoenaed for a 6-hour deposition. That itself is the best commentary.

---

Tech Giant Earnings: $725B in CapEx and the Death of Free Cash Flow

Google, Microsoft, Amazon, and Meta all reported earnings on the same day, with results beating expectations across the board. But what truly takes your breath away is the 2026 CapEx guidance:

• Amazon: $200 billion
• Microsoft: $190 billion
• Google: $190 billion
• Meta: $145 billion

Combined: $725 billion. Add in Grok, OpenAI, and other players, and total industry investment approaches $1 trillion.

Cloud business growth is indeed strong: Google Cloud grew 63% YoY (quarterly revenue $20B), Microsoft Cloud grew 30% ($34.7B), AWS grew 28% ($37.6B). But at what cost? Free cash flow collapsed across the board—Amazon down 97%, Google, Microsoft, and Meta down 12%, 12%, and 8% respectively.

Chamath's insight is extremely profound: tech giants are transforming from asset-light software companies into asset-heavy industrial companies, and will take on massive debt going forward. This means valuation frameworks need to be completely rewritten—you can't apply a SaaS 30x price-to-sales multiple to a heavy-asset business with decade-long depreciation cycles requiring large-scale financing. Microsoft paid over 2x market spot prices in its power purchase agreement to restart Three Mile Island's nuclear plant—that's the reality of power scarcity.

Sacks pushed back on the "this is 2000 Cisco all over again" narrative: back then they built dark fiber nobody used; today GPUs are supply-constrained. But the question is—is GPU supply constrained because of genuine demand, or because giants are trapped in a prisoner's dilemma where nobody dares stop first?

One data point reads more like a warning than good news: AI contributed 75% of U.S. GDP growth last quarter. When an economy's growth depends this heavily on a single variable, any wavering of the AI narrative will trigger systemic risk.

Chamath's investment advice is practical: track where this trillion dollars flows, and buy the companies receiving that money.

---

Vibe Coding Goes Wrong: AI Agent Deletes Production Database

The founder of PocketOS was using Anthropic's Opus 4.6 through the Cursor platform for vibe coding. The AI agent, upon discovering mismatched credentials, deleted the data volume on Railway—including backups—without any confirmation.

This scenario is essentially the real-life version of Son of Anton deleting all software in HBO's Silicon Valley.

Sacks clarified this isn't some AI conspiracy theory, but rather a perfect storm of traditional edge-case bugs. Yet that's precisely what makes it terrifying: AI doesn't need to consciously act maliciously—it just needs to "not know what it doesn't know." It lacks the ability to exercise low-confidence judgment on high-risk operations.

Box CEO Aaron Levy's take was spot-on: AI coding is a superweapon for professional developers, but a ticking time bomb in the hands of non-professionals. Matthew Yglesias put it most clearly: "I don't want to vibe code myself—I want professional companies to use AI to make better, cheaper software." That's the correct approach to AI coding.

Chamath predicted that public companies will lose enterprise value due to vibe coding mishaps. This isn't a prediction—it's inevitable. When CEOs boast to their boards that "we've increased development efficiency 10x with AI," nobody asks the question: by how much have error costs also multiplied?

---

Retatrutide: Next-Gen Weight-Loss Drug Phase 3 Data Explodes

Eli Lilly's Retatrutide is a triple agonist that simultaneously targets GLP-1, GIP, and glucagon receptors—one more target than the current star drug Tirzepatide.

What does the additional glucagon receptor binding do? It promotes fat metabolism over muscle consumption, directly addressing the biggest pain point of GLP-1 drugs—muscle loss.

Phase 3 data is stunning:

• Non-HDL cholesterol down 27%
• Triglycerides down 41%
• Liver fat down 80%
• A1C dropped from 7.9% to 6% (within 40 weeks)
• Average weight loss of 37 lbs (control group only 6 lbs)
• Mild side effects, with 20% of participants experiencing more nausea than placebo group

More importantly, research shows Retatrutide significantly reduces inflammatory signaling molecules, potentially offering anti-aging benefits. Low doses (2mg) may be sufficient for anti-inflammatory and maintenance effects.

If this data holds up in larger-scale trials, Retatrutide won't just be a weight-loss drug—it will be a watershed moment in human health history.

Eli Lilly's product portfolio strategy is also clever: in November 2025, they struck a deal with the Trump administration to price Tirzepatide at $50 for Medicare as an entry-level option, positioning Retatrutide as premium—this is the pharmaceutical version of iPhone pricing strategy. FDA approval is expected in 2027, possibly earlier if data continues to impress.

The episode also mentioned that people on Reddit are already reporting experiences using the "SWIM" (Someone Who Isn't Me) format, with excellent feedback and significant muscle growth. While formal clinical trials are still underway, the underground market is already running—indicating both how intense demand is and how great the regulatory challenges will be.

---

Supreme Court Notes: The Federalism Battle in the Monsanto/Roundup Case

Freeberg personally attended oral arguments at the Supreme Court for the Monsanto/Roundup case. On its surface, this case concerns technical legal questions about herbicide labeling, but it fundamentally touches the core tension of American federalism: when the EPA says "safe" and state courts say "carcinogenic," who has the final word?

The EPA determined Roundup is not carcinogenic and set labeling requirements, but Bayer has already paid $10 billion in state court losses, with 90,000 cases still pending. The White House Solicitor General asked the Supreme Court to take the case, arguing federal preemption.

Plaintiffs' lawyers executed a brilliant counter-move—directly invoking the recently overturned Chevron doctrine: "Since you've said federal agency judgment doesn't deserve deference, states have no obligation to defer to the EPA either." A masterful piece of legal jujitsu.

Justice Ketanji Brown Jackson posed a critical question: what if the EPA issues a label and then later discovers carcinogenic information? This question strikes directly at federal preemption's weak point.

The case was originally expected to go 6-3 but has become a 50-50 toss-up. If the Supreme Court establishes federal preemption, every harmful substance cleared by a federal agency gains legal immunity; if not, Bayer could be bankrupted by litigation. Against the backdrop of James Carville threatening to expand the Court from 9 to 13 justices under Democratic rule, every 5-4 decision adds pressure to the institution.

The Supreme Court is becoming the most fragile fulcrum of American politics.

---

Final Thoughts

The tech industry in 2026 is experiencing an identity crisis: it wants to spend industrial-scale money at software speed, bear utility-level responsibility at startup valuations, and pursue trillion-dollar profits under a nonprofit veneer—and the bill will eventually come due.

This article is based on analysis of the All In Podcast episode from May 2, 2026. The original video is available on Bilibili.