To The Moon

Here's a comprehensive, bullish counter-article based on the title "Beyond the Depreciation Debate: How AI Capex Delivers Real ROI That Burry's Missing". It directly tackles Burry's criticism of extended GPU/server depreciation schedules (which he claims understate expenses by ~$176 billion from 2026–2028, artificially inflating earnings by up to 20%+ for names like Meta and Oracle), frames it as a rational response to real-world tech evolution rather than "fraud," and shifts focus to tangible, accelerating ROI signals from hyperscaler investments in 2026.Beyond the Depreciation Debate: How AI Capex Delivers Real ROI That Burry's MissingMichael Burry's bearish thesis on AI has centered on aggressive accounting at hyperscalers: stretching depreciation lives on Nvidia-powered GPUs and servers from a "realistic" 2–3 years to 5–6+ years. He argues this move—amid massive capex ramps—masks weak returns on investment, sets up an earnings cliff, and echoes dot-com excesses. As of early 2026, his Palantir puts have delivered gains (up ~35% from Q3 2025 entry amid PLTR volatility), while Nvidia puts linger underwater.

finance.yahoo.com

But fixating on depreciation schedules misses the bigger picture. AI infrastructure spending isn't blind speculation—it's delivering measurable productivity gains, revenue acceleration, and economic value today. Hyperscalers aren't extending lives out of desperation; they're responding to rapid utilization ramps, efficiency improvements from newer architectures (like Blackwell), and an inflection toward agentic AI and enterprise deployment. Here's why Burry's depreciation debate underplays the real ROI unfolding in 2026.Depreciation: Accounting Realism, Not Sleight-of-HandBurry contends that Nvidia chips, with 12–18 month product cycles and rapid obsolescence risks, shouldn't see useful lives extended so dramatically. This understates depreciation expense, boosting reported earnings across Big Tech by an estimated $176 billion through 2028 (e.g., potentially overstating Meta earnings by ~21% and Oracle by ~27%).Counterpoint: Useful economic life isn't dictated solely by silicon generations—it's driven by actual deployment data. Nvidia's H100s and early Blackwell systems are seeing high utilization in training and inference workloads. Newer chips deliver massive efficiency leaps (e.g., Blackwell offering ~25x better energy efficiency for inference in some comparisons), allowing older hardware to complement rather than instantly obsolete. Hyperscalers report sustained demand and monetization of capacity "as fast as we can install it."In Nvidia's fiscal Q4 2026 results, data center revenue hit records with 73%+ YoY growth, and Q1 FY2027 guidance came in at ~$78 billion (beating consensus by over $5 billion). Supply commitments nearly doubled to ~$95 billion, locking in capacity through 2027. This isn't circular hype—it's customers (hyperscalers) committing capital because they're seeing returns in ad optimization, cloud workloads, coding assistants, and emerging agentic systems. Extended depreciation reflects this durability, not manipulation. Burry's strict 2–3 year assumption feels conservative when real-world inference economics and multi-year platform value are factored in.The $700 Billion Capex Bet: Not Blind Spending, But Strategic BuildoutBurry questions whether hyperscaler capex (projected near $700 billion combined for Alphabet, Amazon, Meta, and Microsoft in 2026 alone) can ever justify itself, especially with front-loaded infrastructure costs.

cnbc.com

Evidence points to yes—ROI is emerging faster than skeptics expected:Monetization signals: Amazon noted adding capacity and monetizing it rapidly. Meta and Google highlighted AI-driven improvements in ad relevance and user engagement. Microsoft tied Azure AI growth to enterprise copilots and productivity tools.

Inference inflection: 2026 marks a shift from training-heavy spend to inference (token generation, agents). Nvidia's Jensen Huang has called this the "agentic AI" phase, where compute generates direct economic value through automation and better decision-making.

Productivity data: Surveys and studies in 2025–2026 show AI delivering task-level gains of ~25–30% in high-skill areas (coding, analysis, customer service). Leading adopters report revenue uplift or cost savings; broader economy-wide TFP (total factor productivity) boosts are projected at 0.1–0.2pp annually, compounding over time. PwC and others note that top-performing companies (focused on growth use cases) are capturing outsized gains—up to 4x productivity acceleration in pockets.

This isn't dot-com 2.0 speculation on unproven consumer apps. It's foundational infrastructure for an AI-native economy: data centers as the "new electricity," enabling compounding advantages in software, robotics, healthcare, and defense. Palantir exemplifies the software layer—its AIP platform turns raw data into deployable agents. In Q4 2025, U.S. commercial revenue surged 137% YoY; full-year 2026 guidance implies ~61% total revenue growth to ~$7.19 billion, with U.S. commercial up at least 115% to over $3.1 billion. Backlog and deal velocity (via AIP bootcamps) show sticky, high-margin adoption—not vaporware.Why ROI Is Real (and Accelerating) in 2026Burry's focus on near-term accounting optics overlooks multi-year flywheels:Hyperscaler economics: Even with heavy 2026 capex, these firms maintain strong balance sheets and cash flows. Many guide for continued operating income growth alongside AI investments. Amazon, for instance, expects power capacity to double again by 2027.

Nvidia's visibility: FY2026 revenue grew 65%+ from a massive base; gross margins held ~75%. Roadmap clarity (Blackwell ramp, Rubin ahead) plus sovereign AI and industrial use cases add diversification beyond pure hyperscalers.

Enterprise proof points: Palantir's "commodity cognition" and ontology-based platforms deliver measurable outcomes for customers in manufacturing, finance, and government—driving Rule of 40 scores well above peers.

Early 2026 volatility (PLTR pullbacks, broader AI digestion) reflects digestion after 2025 euphoria and macro noise, not fundamental failure. Nvidia's puts staying underwater reflects resilient demand; capex isn't peaking—it's transitioning to higher-ROI phases.Risks Remain—But the Thesis HoldsNo one denies execution risks: competition (AMD, custom chips), energy constraints, or a macro slowdown could pressure timelines. Valuation multiples on AI leaders demand near-perfect execution. Burry's track record demands respect—his shorts have profited on volatility.Yet the structural case is stronger: AI capex is infrastructure spend with network effects, data moats, and defensibility (Nvidia's CUDA, Palantir's platforms). Historical parallels (cloud computing in the 2010s) show initial "overinvestment" phases yielding massive long-term returns. Burry's depreciation critique highlights a legitimate accounting debate, but it doesn't negate the underlying value creation.The real risk for bears may be missing the acceleration: 2026 isn't the peak of AI hype—it's when ROI starts proving the spend was worth it. Hyperscalers and enablers like Nvidia and Palantir are building the rails for an AI economy that compounds for decades.Burry may win rounds on timing and volatility, but the long-term ROI story for AI capex is just getting started.This is not financial advice. Do your own research. AI investments involve significant risks, including technological shifts, competition, and macroeconomic factors.