Zitron: "Everyone Has Been Sold a Lie" on AI

AI investment “narrative vs reality” (core claim)

The speaker argues investors are being misled about where AI spending (especially in cloud/datacenters) ultimately goes—specifically, that heavy capex by hyperscalers is being used to build/serve infrastructure for two largely unprofitable AI model companies rather than producing broad, durable ROI.

Tickers mentioned (companies/instruments):

• Microsoft (MSFT)
• Google (Alphabet / “Alphabet”)
• Amazon (AWS / “Amazon”)

Counterparty / infrastructure and partner references:

• OpenAI (not publicly traded in the excerpt, but discussed)
• Anthropic
• SoftBank (shareholder/customer references)
• Broadcom (mentioned as selling TPUs to Google)

---

Revenue exposure estimates (Google Cloud dependence on OpenAI/Anthropic)

UBS estimates (as stated)

• 27% of Google Cloud revenue this year comes from OpenAI + Anthropic
• Increasing to over 48% next year
• Described as “over $124 billion next year” in revenue tied to that exposure

Clarification noted by the speaker: the UBS figures refer to both OpenAI and Anthropic together.

Barclays estimates (as stated)

• 13% of all revenue this year from OpenAI + “Strike” (auto-subtitle ambiguity likely; context suggests OpenAI/Anthropic-related demand)
• 18% next year, described as a bigger business than Google Cloud

Supporting point

• OpenAI is described as a large customer of Google Cloud, implying most investors underappreciate the concentration.

---

Materiality of OpenAI to Microsoft growth (2025)

The speaker claims (via “own reporting”):

• In 2025, 69% of the year-over-year growth of Microsoft’s Intelligent Cloud segment came from OpenAI.
• Without OpenAI, growth would be only ~8% YoY, described as “barely beating inflation.”

---

“Party ends” catalysts: IPO timing + funding dependence

OpenAI IPO timing

• Potential flashpoint: OpenAI IPO is described as delayed.
• Reported consideration (via New York Times) to push from the expected window to 2027.

Funding/cash-flow risk

• OpenAI and Anthropic are characterized as needing continual capital because they “do not pay their bills out of existing cash flow.”
• If capital becomes scarce, the speaker frames this as an initial “domino” risk.

---

Data center buildout lag (macro/infrastructure timing risk)

Build timeline

• Data center capacity build timeline cited: 12 to 36 months, depending on launch size.

Throughput/demand math example

• Sightline Climate estimate: about 190 GW of data-center capacity planned/built in coming years.
• Assuming PUE (efficiency) = 1.3:
  • The excerpt implies extremely large required energy/throughput demand (subtitle phrasing appears inconsistent/likely an auto-error).
• Claimed revenue need:
  • Over $1.6 trillion of annual revenue required to “satiate” that capacity.

Concentration problem (speaker’s conclusion)

• Serving that demand with only two major customers (OpenAI + Anthropic) “won’t” work unless they grow massively.

---

Balance sheet / financial stress claims (OpenAI)

The speaker cites (via “audited financials” reported in the Financial Times):

• OpenAI allegedly lost $20.9 billion in 2025.

Additional claims:

• Over $800 million of OpenAI revenue came from SoftBank related to a “Crystal intelligence” program, with the speaker saying they found no evidence of actual activity.
• Shareholder/control angle: SoftBank described as a large shareholder of OpenAI with no board seats (as stated).

---

Infrastructure “circular financing” + double-revenue mechanism (TPUs)

The speaker argues hyperscalers create demand loops:

• They build infrastructure for AI labs and also position themselves as suppliers of compute capacity.

Google example described:

• Google obtains TPUs from Broadcom
• Those TPUs are used/sold to Anthropic
• Then rented back to Anthropic “through Google,” so Google “double up on revenue.”

Caution (speaker’s view):

• It may look attractive, but the speaker argues it’s unsustainable because Anthropic and OpenAI are unprofitable and the ROI isn’t clear.

---

“Compute cost” and dependence on massive scale

Funding and infrastructure cost claims

• OpenAI and Anthropic collectively raised $200–$300 billion (as stated).
• The speaker claims they effectively received infrastructure buildup funded/paid by Microsoft, Google, Amazon, reducing direct need to pay.
• A Microsoft executive (mentioned via trial context) is quoted as saying the infrastructure “cost $100 billion.”
• The speaker translates that to roughly $70–$80 billion.

Barrier to entry

• Conclusion: the scale creates barriers to entry (“nobody else can get as big” / “no one else could afford that compute”).
• “China is coming up behind them” is mentioned as competitive pressure (not quantified).

---

What happens to hyperscaler stocks if ROI fails?

The speaker frames scenario risk tied to labor productivity:

• If AI delivers productivity gains and requires fewer workers, that supports the market’s “hope” for ROI.
• If it doesn’t happen, they question how the market valuation for Amazon/Google/Microsoft unravels.

Business model description

• Hyperscalers are described as transitioning into GPU-filled “asset mongers” with semi-built datacenters.
• They are said to rent these assets to serve a small number of customers (2–3 at best) to rent out models.
• “Not selling compute capacity yet” is mentioned (monetization clarity appears unclear in the subtitles).

Explicit valuation/demand mismatch claim

• UBS estimate cited again:
  • ~$76 billion in 2027 of Google Cloud revenue coming from Anthropic.
• Speaker’s argument:
  • Anthropic reportedly burned tens of billions, so it may not afford such future payments.
• Final caution:
  • For the thesis to hold, OpenAI/Anthropic must grow enormous for datacenter capacity to be “good” (i.e., justified by demand and spend).

---

Methodology / framework mentioned (implied)

• Concentration/exposure approach: estimate the % of cloud revenue attributable to specific AI labs (UBS/Barclays), then infer survivability/negotiating leverage.
• Customer capacity utilization math: use planned data-center capacity (GW), PUE, and implied revenue scale to test whether only two customers can absorb demand.
• Segment contribution attribution: estimate portion of Microsoft Intelligent Cloud YoY growth attributable to OpenAI.
• Scenario risk analysis: “party ends” triggers (e.g., IPO delay to 2027, capital constraints) plus ROI failure tests (productivity not materializing → valuation repricing).

---

Key numbers and timelines (as stated)

• Cloud revenue concentration (Google Cloud):

  • 27% (this year) → >48% (next year) of revenue from OpenAI + Anthropic (UBS)
  • “> $124B next year” tied to that exposure (UBS, as stated)

• Revenue impact / growth attribution (Microsoft Intelligent Cloud):

  • 69% of YoY growth (2025) attributed to OpenAI
  • Without OpenAI: ~8% YoY growth

• Buildout timeline: 12 to 36 months

• Capacity/revenue scale:

  • 190 GW planned/built (Sightline Climate, as stated)
  • PUE = 1.3 (as used)
  • Implied required revenue: over $1.6 trillion annually (as stated)

• Financial stress claims:

  • OpenAI losses: $20.9B in 2025
  • OpenAI revenue from SoftBank: >$800M (Crystal intelligence, as stated)

• IPO / “flashpoint”:

  • OpenAI IPO potentially delayed until 2027

• Funding/infrastructure cost:

  • OpenAI + Anthropic funding: $200–$300B (as stated)
  • Infrastructure cost referenced: $100B (exec quote), speaker estimate $70–$80B

• UBS estimate (Anthropic monetization):

  • ~$76B in 2027 Google Cloud revenue from Anthropic (as stated)

• Contractability risk:

  • Anthropic reportedly burned tens of billions (no specific number provided beyond that)

---

Disclosures / disclaimers

• No explicit “not financial advice” disclaimer appears in the subtitles provided.

---

Presenters / sources mentioned (end)

• Ed Zitron (also rendered as “Ed Citron”) — CEO, Easy Primary Research (mentioned as the speaker)
• Stephen Ju — referenced as the UBS source (“UBS estimates … mentioned by Stephen Ju”)
• Dario Amodei / Dario Ahmadi of Anthropic — mentioned in the context of prior remarks about productivity (name appears as “Dario Ahmadi” in subtitles)
• UBS — source for cloud revenue exposure estimates
• Barclays — source for alternate revenue/exposure estimates
• Sightline Climate — source for data-center capacity estimate
• New York Times — reported on OpenAI IPO delay to 2027
• Financial Times — referenced as where audited financials for OpenAI were reported