SoftBank Sells Into Strength, Burry Questions the Math—Why These Two Threads Matter Now

Two storylines have dominated market chat in early November: first, SoftBank’s decision to sell down its Nvidia stake—roughly $5.8 billion at October prices—after a historic run; second, Michael Burry’s critique that parts of Big Tech’s reported profits are flattered by lengthening the ‘useful life’ of AI infrastructure, which mechanically reduces depreciation expense and smooths earnings through the upcycle. The timing is not incidental: these developments arrive just weeks after Nvidia briefly touched a $5 trillion market cap in late October, a milestone that crystallizes how concentrated the AI trade has become.

Let’s separate heat from light. SoftBank trimming winners is not new—this is the same house that took early AI platform risk via Arm, rotated aggressively between venture assets, and now monetizes liquidity when it’s abundant. It’s entirely consistent with Masayoshi Son’s stated ambition to recycle gains into the next platform bets (including AI-native chips and robotics) rather than passively ride a single ticker. The fact set is straightforward: SoftBank sold a stake in Nvidia in October valued at about $5.8 billion, locked in profits after a parabolic move, and did so during a window where the whole market was re-basing AI expectations around power, packaging, and model-economics constraints.

On a separate axis, Burry’s criticism is not about whether AI matters; it’s about how the profits are presented. His public remarks singled out how certain companies in the AI stack (he mentioned Meta and Oracle among others) have extended the useful life of servers and other capital equipment. Under both US GAAP and IFRS, increasing useful life reduces current-period depreciation, lifting operating income and EPS without changing cash reality. In other words, if the cash leaving the door for capex remains high while the expense recognized each quarter falls, you can generate the appearance of expanding margins. Burry’s message: don’t confuse optics with economics. ([Yahoo Tài Chính][1])

Context Check: What the Tape Has Already Told Us

To avoid being whipsawed by the latest headline, anchor on what the tape already priced: AI leadership has compounded into a handful of mega-cap nodes—training silicon (Nvidia), hyperscaler capex pipelines, and a narrower layer of winners in memory, substrates, and EDA. By late October, Nvidia had—per multiple mainstream reports—briefly traded above a $5T valuation. Even if the exact level oscillates week to week, the directional truth is unambiguous: a single vendor’s equity value now rivals the GDP of large G20 nations and represents a material percentage of the S&P’s total free float. That means two things for portfolio math: index beta is now AI beta, and small changes to AI capex outlooks will propagate through passive vehicles with surprising violence.

Is that automatically a bubble? Not necessarily. Remember, trains that look like bubbles from the outside can be cash machines on the inside if three conditions hold: (1) demand is recurring rather than one-off (model retraining, inference growth, agentic workloads); (2) supply bottlenecks (power, packaging, HBM) keep pricing and utilization high enough to protect margins; and (3) downstream software and services monetize the compute into durable revenue per user (RPU). That trifecta converts headline hype into durable fundamentals.

Why SoftBank’s Trim Is More Signal Than Shade

SoftBank cuts positions for portfolio architecture reasons: one part liquidity management, one part optionality for the next turn. The October sale into Nvidia’s strength—reported around $5.8B—is consistent with a long-standing playbook of crystallizing gains when a position dominates risk budget and recycling capital into earlier-stage convexity (think Arm, robotics, and AI-native chips). Crucially, nothing in the SoftBank disclosure suggests a thesis break for AI itself; if anything, their core exposure (via Arm and private holdings) remains tightly coupled to the AI compute explosion. Read the sale as a rebalance, not a repudiation.

Still, when a high-profile sponsor sells, it sharpens the debate around marginal buyers. In 2023–2025, the marginal buyer of AI leaders has often been the same cohort: US megacaps allocating capex, sovereigns building compute footprints, and passive funds pulled by index mechanics. If the marginal buyer’s pace slows—because power constraints delay data center buildouts, because HBM supply is rationed, or because governance forces hyperscalers to stagger spend—the multiple that the market is willing to pay for today’s growth can compress even as absolute profits remain large. That is how you get the paradox of ‘great companies, falling stocks.’ SoftBank’s trim doesn’t cause that; it illuminates it.

Burry’s Depreciation Thesis in Plain English

Depreciation is the accountant’s way to spread the cost of long-lived equipment over its life. If a firm buys $10B of servers and assumes they last 3 years, it books roughly $3.3B of depreciation per year (ignore salvage for simplicity). If, after a policy change, the firm says those servers last 5 years, annual depreciation falls to $2B. That $1.3B difference is not free money; it is an optical boost to operating profit. Cash capex still happened. Free cash flow is constrained by cash outlays (and future replacement cycles), not by the non-cash depreciation route through the income statement.

Burry’s contention is that the AI buildout—massive, front-loaded, and hardware-centric—invites exactly this kind of smoothing. Extend useful life here, broaden capitalization policies there, and your GAAP earnings glide path looks friendlier while the economic reality (power contracts, lease obligations, replenishment capex) remains demanding. His public remarks singled out specific companies and suggested that, across the complex, headline EPS may overstate ‘core profitability’ at this early stage of the capex supercycle. Whether you agree or not, the test he implies is simple: track cash. The earnings you can bank are the ones that clear cash reinvestment + cash taxes + cash interest.

How to Reconcile ‘Record Highs’ With ‘Accounting Worries’

Markets are discounting machines. If the slope of AI unit economics bends positive—because utilization stays high, because inference workloads explode beyond chat into agents and automation, because power constraints create temporary scarcity premiums—then today’s valuations can be justified even with some accounting smoothing at the edges. Conversely, if demand breathes while supply catches up (new fabs, new HBM lines, accelerating packaging) and if CFOs everywhere are stretching useful lives, the aggregate lift to EPS can mask a downshift in cash returns. That is not criminal; it is cyclical. But it means your screening tools must privilege cash conversion and maintenance capex realism over headline EPS beats.

Three Things the AI Trade Must Get Right From Here

1. Energy & Power Availability. The physics is not optional: data centers hungry for 100–500MW class projects face interconnection queues, transformer lead times, and local politics. If the grid lags, the demand curve could be forced to flatten temporarily, pushing some buyers to defer orders. In that world, stocks with the most shipment-timing risk will rerate first.
2. Packaging & Memory. HBM supply and advanced packaging (CoWoS and equivalents) remain gating items. If bottlenecks persist, price/mix can stay unusually strong, supporting margins. If bottlenecks ease quickly, price discipline will matter more than unit volume for sustaining profits.
3. Demand Broadening. For the cycle to be durable, spend must broaden beyond hyperscalers into enterprise verticals (healthcare, finance, industrials) with ROI-positive workloads. The ‘agentic turn’—LLM-driven software agents actually completing tasks—will be crucial to move from trials to production.

Is the Market Already Too Concentrated?

Concentration is not a thesis; it is a measurement. The practical risk is that passive flows magnify drawdowns by forcing the same hands to sell the same names when factor signals flip. That does not invalidate AI; it changes the path. In concentrated regimes, portfolio construction becomes as important as stock-picking: balance crown-jewel compute with picks-and-shovels (power equipment, substrates, thermal), with royalty-like models (EDA, IP), and with through-cycle beneficiaries (utilities, grid services). If the crown jewels wobble on a headline or a policy tweak, the shovels should keep printing.

Signal Versus Noise: What We Actually Learned From SoftBank

We did not learn that ‘AI is over.’ We learned that professional liquidity managers harvest gains when multiples get saharan and when their own risk budget concentrates. We also learned that big buyers exist even at $5T—SoftBank needs counterparties to sell to. The marginal buyer is there; the question is at what price, with what hurdle rate. If power constraints or budgeting cycles defer some 2026 deliveries into 2027, your expected IRRs shift—not because the end-state shrank, but because the path zig-zagged. That’s how cycles breathe.

How to Adjust Your Playbook If Burry Is Right (Even Partly)

• Normalize the income statement. Recast earnings by resetting useful lives to pre-change assumptions and expensing a reasonable ‘maintenance capex’ proxy—e.g., the rolling 3-year average of replacement spend needed to hold capacity constant at target utilization. Compare that to reported EBIT and FCF.
• Follow cash taxes and interest, not just EBIT. If earnings are ‘high quality,’ cash taxes will climb and cash interest coverage will improve without borrowing more.
• Watch prepayments and backlog quality. Prepaid commitments from customers (or outsized backlog disclosure) are excellent tells for durability. If prepayments fade while reported margins remain buoyant, the cash/earnings gap is widening.
• Lean into second-derivative winners. Grid interconnectors, transformer makers, HBM players, substrates, and cooling provide exposure to AI capacity without the same accounting debates over server life.

But What About That ‘$500B in 48 Hours’?

Hyperbolic or not, the spirit of that claim lands: index-scale cap adds and wipes have become normal around AI catalysts. The healthier interpretation is not ‘ignore it’ but ‘plan for it.’ Volatility clusters where narrative density is highest. If you model position sizing as a function of power capacity, packaging throughput, and prepay trends—not just quarterly beats—you can hold winners through the chop without outsourcing conviction to headline writers.

Where This Cycle Could Break (and How It Could Get Even Better)

Break risks: power buildouts slip, HBM/packaging constraints ease abruptly and flip bargaining power to buyers, AI agent productivity fails to materialize at scale, and governments tighten antitrust around AI platform bundling. Upside risks: autonomous agents move from demo to factory floor, inference spend ramps in consumer and enterprise simultaneously, and grid modernization outpaces expectations, keeping utilization and pricing elevated. In the upside, the duration of the AI capex wave stretches—and with it, the time firms have to grow into valuations.

Bottom Line

SoftBank selling a multibillion-dollar Nvidia stake is a portfolio decision, not a funeral dirge for AI. Burry’s accounting critique is a measurement challenge, not a refutation of the technology’s transformative arc. Both are useful precisely because they force investors to quantify the difference between optical earnings and cash compounding. If you insist on cash-first models, normalize for depreciation games, and size positions to the real-world bottlenecks (power, packaging, memory), you don’t need to guess whether this is a bubble; you can endure a rerate and still own the future.

Notes & sources: SoftBank’s October sale of its Nvidia stake (circa $5.8B) and contemporaneous context are drawn from recent reporting by The Guardian; Nvidia’s late-October brush with $5T is also covered there. Burry’s accounting critiques and references to useful-life extensions have been reported across financial media, including Morningstar/MarketWatch roundups and The Economic Times.