Using Claude.AI free subscription with the following prompt “Read this text and point out the hidden assumptions, biases, or unspoken perceptions that most readers would overlook – but experts would notice.” to review the BOND AI Trends Report (May 2025).
Meta-Level Observations Experts Would Notice
1. Venture Capital Framing Bias
The entire document is written from a VC perspective optimizing for growth-at-all-costs narratives rather than sustainable business models.
Hidden assumptions:
- “High cash burn + high growth = inevitable success” (pages 180-181: Amazon, Tesla, Uber examples)
- Survivorship bias: Only discusses companies that survived massive losses, not the hundreds that burned out
- The framing “it’s different this time” (p.154) is the exact phrase that precedes most bubble bursts, yet presented as potentially valid
What’s unspoken:
- BOND has portfolio positions in many companies discussed (disclosed on p.339 but not integrated into analysis)
- The report needs readers to believe in continued AI spending to maintain portfolio valuations
- No discussion of what happens if the monetization never materializes
2. The “Inevitability” Narrative
Embedded assumption: AI adoption is inevitable and uniformly positive
Evidence:
- Page 7: “AI usage is surging” (no discussion of plateau risks)
- Page 21: “AI is a compounder” (assumes compounding continues indefinitely)
- Page 326-327: Shopify & Duolingo memos presented as wisdom rather than corporate mandates
What’s hidden:
- Technology adoption curves often have inflection points where growth stalls
- No discussion of “AI fatigue” or declining marginal utility
- Assumes linear improvement when many technologies hit plateaus (see: Moore’s Law slowing)
Expert observation: This is classic technological determinism – the belief that technology drives change rather than being shaped by social/economic forces.
3. The China Threat Amplification
Pages 271-298: Extensive focus on China’s AI capabilities
Hidden assumptions:
- Zero-sum thinking: China’s success = America’s failure
- Geopolitical framing benefits the thesis: If AI is a “space race,” then infinite spending is justified
- Data reliability issues buried: Multiple footnotes say “China data may be subject to informational limitations” but charts treat data as equally reliable
What’s unspoken:
- This framing benefits defense tech companies (Anduril, Palantir) that BOND likely invests in
- The “threat” narrative justifies higher valuations for US companies
- No discussion of collaboration potential or open-source as reducing geopolitical risk
Expert catch: Page 271 quote from Meta’s CTO about the “space race” is presented as objective analysis, but it’s a stakeholder comment from someone whose company benefits from this framing.
4. Revenue Multiples Are Justified by Historical Precedent
Pages 178-179: OpenAI valued at 33x revenue vs. median 6.9x
Hidden assumption: Historical loss-making companies (Amazon, Tesla) prove that today’s burn rates are acceptable
What’s missing:
- Interest rate environment: Amazon/Tesla burned during near-zero rates; today’s cost of capital is 4-5%+
- Market conditions: Those companies had less competition and clearer paths to dominance
- Scale differences: OpenAI’s $5B loss on $3.7B revenue is a 135% burn rate – historically unprecedented at this scale
- No discussion of: What if inference costs fall faster than they can build moats?
Expert observation: The comparison is apples-to-oranges. Amazon had network effects and physical infrastructure moats. AI models face commoditization (openly acknowledged on p.142).
5. The “Picks and Shovels” Misdirection
Pages 108-109: NVIDIA presented as the clear winner
Hidden assumption: Infrastructure providers are safe bets because they win regardless of which applications succeed
What’s unspoken:
- Custom silicon threat buried: Pages 162-163 show Google/Amazon building their own chips (TPUs, Trainium) with better price/performance
- NVIDIA’s lock-in is weakening: Open-source models + commoditized hardware = pressure on margins
- Historical precedent ignored: Cisco was the “picks and shovels” of Internet 1.0, peaked in 2000, took 24 years to regain that market cap
Expert catch: The report shows NVIDIA at 25% of global data center CapEx (p.109) but doesn’t ask: “What happens when that falls to 15% due to ASICs?” This is the single biggest risk not adequately explored.
6. The Employment Optimism
Pages 324-336: “Technology creates more jobs than it destroys”
Hidden assumptions:
- Past = Future: “Labor productivity +31% since 2000, employment +89%” (p.335) assumes this continues
- Transition speed ignored: Previous transitions took decades; AI is moving in quarters
- Job quality unexamined: Doesn’t distinguish between high-skill and gig economy jobs
What’s unspoken:
- Page 324 admits: “this time it’s happening faster” but doesn’t explore implications
- The entire section avoids discussing wage pressure or inequality
- Scale AI (p.170) revenue model is literally “humans training AI” – a temporary business model
Expert observation: The quote from NVIDIA’s Jensen Huang (p.336) – “you won’t lose your job to AI, but to someone using AI” – is corporate speak for “we’re enabling mass workforce reduction but deflecting responsibility.”
7. The Open-Source Paradox
Pages 261-269: Open-source models closing the gap
This is the most buried lede in the entire document:
- Page 264: DeepSeek R1 achieved 93% vs. OpenAI’s 95% on math tests
- Page 265: DeepSeek achieved parity with 96% lower training costs (chart shows $5M vs. $100M+)
- Page 268: Meta Llama downloads +3.4x in 8 months to 1.2B
What this means (unstated):
- If open-source reaches parity, all the frontier model companies lose pricing power
- The entire “$95B raised vs. $11B revenue” dynamic (p.176) collapses
- Winner-take-all assumptions break down
Why it’s buried:
- Acknowledging this undermines the entire investment thesis for closed-model companies
- The report presents it as “healthy competition” rather than “existential threat to monetization”
Expert read: This is the single most important trend in the document, but it’s in the “threats” section rather than front-and-center. This is editorial framing to minimize cognitive dissonance.
8. The Energy Bottleneck Handwaving
Pages 124-128: Data centers consuming 1.5% of global electricity, growing 12%/year
Hidden assumptions:
- Energy constraints are solvable with investment
- Grid capacity can scale linearly with demand
- NVIDIA’s efficiency gains (105,000x improvement, p.136) will offset usage growth (Jevons Paradox acknowledged but then ignored)
What’s unspoken:
- Political risk: No discussion of regulatory backlash to AI energy consumption
- Competition for resources: AI data centers vs. EVs, heat pumps, manufacturing
- Build times: Power infrastructure takes 5-10 years to deploy; xAI’s 122-day data center (p.122) is an extreme outlier, not scalable
Expert catch: The report celebrates xAI building a data center in 122 days (vs. 234 for a house) but doesn’t ask: “Where did the electrical substation come from?” Those take years and are the actual bottleneck.
9. The Vertical SaaS Escape Hatch
Pages 214-243: Vertical AI software growing faster than horizontal platforms
This section is strategically positioned to answer the “but what if model margins collapse?” question
Hidden assumptions:
- Vertical AI will be defensible where horizontal AI is not
- First-movers in verticals (Harvey for legal, Abridge for healthcare) will maintain leads
- Incumbents won’t just add AI features (even though the report shows they’re doing exactly that)
What’s missing:
- Why wouldn’t ChatGPT Enterprise just add legal/healthcare modes? (Page 228 shows they have 20M business users)
- Data moats aren’t obvious: Most “vertical” data is unstructured text – trainable by general models
- Switching costs are low: These are mostly thin wrappers on foundation models
Expert observation: This section reads like portfolio company pitch decks rather than objective analysis. The growth rates are impressive (Cursor $1M→$300M ARR in 25 months, p.233) but unit economics and defensibility are never examined.
10. The “2.6B New Users” Hail Mary
Pages 309-322: New internet users will be “AI-native”
This is the most speculative section, presented as high-conviction
Hidden assumptions:
- SpaceX Starlink will connect the unconnected (5M subscribers vs. 2.6B target = 0.2% penetration)
- These users will monetize despite being in low-GDP regions
- AI-first interfaces will be better for new users (unproven)
What’s unspoken:
- ARPU will be low: India has 14% of ChatGPT users but lower income (p.316)
- Infrastructure costs are high: Satellite internet is expensive to operate
- This is a 10-year bet presented as near-term catalyst
Expert read: This section exists to counter the “market saturation” concern. It’s saying “even if developed markets slow, there’s 2.6B more users coming.” But the unit economics don’t work – these are the users least able to pay $20/month for ChatGPT Plus.
Stylistic & Structural Biases
11. Cherry-Picked Comparisons
Throughout: Comparisons are always to the best-case historical precedents
Examples:
- ChatGPT compared to Google (winner) not to Clubhouse (flamed out)
- AI adoption compared to Internet 1.0 (success) not to 3D TV or QR codes (failed despite hype)
- China AI compared to Sputnik (motivated USA) not to Japan 1980s tech panic (overblown)
What’s missing: Any discussion of failed technology waves or bubbles that burst
12. Metric Selection Bias
What’s measured:
- Revenue growth rates (always impressive for startups)
- User growth (vanity metric without retention/engagement depth)
- CapEx spending (presented as strength not risk)
- Model performance on benchmarks (ignoring real-world utility gaps)
What’s NOT measured:
- Customer acquisition costs
- Net retention rates beyond year 1
- Gross margins for model providers (buried: OpenAI losing money on every query)
- Developer churn rates (e.g., how many stop using Cursor after trying it?)
Expert observation: This is classic growth-stage VC metrics – optimized to show momentum, not sustainability.
13. The Timing Sleight-of-Hand
Throughout: Growth rates calculated over cherry-picked time periods
Examples:
- “ChatGPT 5.5x faster than Google to 365B searches” (p.20) – compares 2 years to 11 years, ignoring that Google had to build internet infrastructure first
- NVIDIA revenue “+28x over ten years” (p.161) – includes crypto boom/bust, so cyclical peaks distort trend
- “AI job postings +448%” (p.332) – measured from 1/18, before ChatGPT existed, so base is artificially low
What this hides: Deceleration. Many of these curves are starting to flatten but the long timeframes mask it.
14. The AGI Bait-and-Switch
Pages 92-93: AGI discussed as “reachable threshold” with Sam Altman quote: “We are now confident we know how to build AGI”
What’s unspoken:
- AGI is undefined – no consensus on what it means
- Altman has incentive to claim progress (fundraising, talent recruitment, regulatory capture)
- The report doesn’t commit to when (“timelines remain uncertain”)
Why it’s included:
- AGI justifies unlimited spending – if we’re close to artificial general intelligence, any price is worth paying
- It reframes current losses as “R&D toward AGI” rather than “unsustainable business model”
Expert read: This is vaporware marketing. Notice how the report immediately pivots to “still, the implications warrant a measured view” (hedging) after floating the AGI balloon.
The Unspoken Meta-Narrative
What the Report Is Really Arguing:
- AI is inevitable and transformative (pages 1-51)
- Adoption is happening faster than any prior technology (52-128)
- Current business models don’t work YET (153-247)
- But historical precedent says patient capital wins (180-181)
- And if USA models don’t win, China will (248-298) ← creates urgency
- Plus there’s huge greenfield opportunity (299-322) ← creates hope
- So you should keep investing (implicit throughout)
What’s NOT Being Said:
“We are in the middle of the largest capital misallocation event in technology history, and it may not end well for most participants.”
Evidence buried in the report:
- Training costs rising faster than inference costs falling (net margin squeeze)
- Open-source closing gap (commoditization)
- Model performance converging (differentiation collapsing)
- Energy becoming bottleneck (growth constraint)
- Profitability timeline undefined (risk to investors)
The Most Important Hidden Assumption
The Entire Report Assumes: “Whoever spends the most, fastest, wins”
This is the unexamined premise underlying everything:
- Big Tech spending $212B/year on CapEx (p.97) = good
- OpenAI burning $5B (p.173) = necessary
- DeepSeek spending 96% less (p.286) = temporary disadvantage
But what if the opposite is true?
What if capital efficiency wins in a world where:
- Inference costs approach zero
- Open-source reaches parity
- Commoditization is inevitable
The report acknowledges this risk obliquely (p.130):
“Training is expensive, serving is getting cheap, and pricing power is slipping.”
But then spends 200+ pages arguing why you should ignore that dynamic.
The Expert’s Summary
This report is: ✅ Comprehensive data compilation ✅ Well-researched growth metrics ✅ Insightful on adoption patterns
But it’s also: ❌ Systematically optimistic on monetization ❌ Dismissive of commoditization risk ❌ Conflating growth with defensibility ❌ Using geopolitical fear to justify valuations ❌ Cherry-picking historical comparisons ❌ Written to support a specific investment thesis
The most sophisticated readers will notice:
The document is structured like a legal brief – marshaling evidence for a predetermined conclusion rather than objectively weighing both sides. Every apparent “concern” (open-source, China, energy) is reframed as “manageable” or “opportunity.”
The question experts are left asking:
“If this is the BEST-CASE analysis from sophisticated VCs with proprietary data… what does the bear case look like?”
Leave a Reply