Bernstein: Core Scientific’s 75% AI return is a capex-skewed outlier, not a blueprint for bitcoin miners

Investors want a clean narrative: miners redeploy power and racks into AI, margins explode, equity re-rates. The problem is that a single exceptional datapoint is being treated like a baseline. Bernstein argues Core Scientific’s 75% return on its AI deal sits in outlier territory, driven by capex advantages that are hard to copy. By contrast, the firm pegs stabilized return on assets at roughly 5% for TeraWulf and 4% for Cipher. That spread is the tell.

The point worth dwelling on is capex-advantaged outperformance. When a miner posts eye-popping returns in AI hosting, it often reflects a peculiar mix: sunk infrastructure from prior cycles, unusually favorable power contracts, and timing that lets legacy buildouts re-price into a hotter demand curve without paying today’s replacement cost. Those ingredients are scarce and perishable.

Why this matters: - Cost stack divergence: Bitcoin mining monetizes hashrate with flexible uptime and limited SLA pressure. AI hosting monetizes sustained, high-density compute with strict power and availability requirements. The same megawatt does not earn the same risk-adjusted return once you price in redundancy, cooling, networking, and service layers. - Replacement cost vs. historical cost: Stabilized ROA anchored to yesterday’s build price will screen high. Rebuild it today—new switchgear, higher rack density, interconnect upgrades, thermal retrofits—and the economic profile compresses quickly. That is where 5–4% stabilized ROA estimates become plausible. - Contract quality over headline IRR: Term length, escalators, pass-throughs on power, and penalties for downtime drive realized returns. One premium contract can dominate a case study; a portfolio of average contracts tends to settle closer to mid-single-digit ROA once financing, maintenance capex, and contingencies are layered in. - Capital structure drift: Equity markets sometimes capitalize “story IRRs” while debt markets underwrite to coverage. If the cost of capital sits above single-digit stabilized ROA, scale can add revenue without creating durable equity value.

There is also a behavioral trap. After the halving, many miners chased alternative revenue per megawatt. AI is the loudest option, so investors extrapolate a 75% print across the sector. That shortcut ignores execution risk: procuring GPUs, meeting density targets, delivering Tier-appropriate uptime, and staffing with DC operations talent. Bitcoin mining is industrial; AI hosting is industrial plus service. Margins move accordingly.

How I would evaluate miner-to-AI pivots: - Normalize for replacement cost: Rebuild the site pro forma at today’s prices and re-run ROA. If it breaks below the company’s blended cost of capital, growth won’t compound value. - Underwrite power like a lender: Volatility, curtailment rights, and basis risk should be explicit in the model. Assume less indexation pass-through than the deck suggests. - Demand credible ramp curves: GPU supply, network buildouts, and customer onboarding rarely follow perfect S-curves. Penalize delays. - Separate bitcoin optionality from AI services: Assign different hurdle rates. Mixing them masks risk. - Track service intensity: SLAs, managed services, and networking support can lift price per kW but also raise opex and liability.

Ethically, steering investors with an outlier case study sets unrealistic expectations for retail buyers who may not parse stabilized vs. initial returns. The right framing acknowledges that some operators will earn premium spreads for a period due to unique positioning, while sector-wide economics likely gravitate toward mid-single-digit ROA until a true scarcity of qualified capacity persists for longer than markets expect.

Bernstein’s take—Core Scientific’s 75% AI return as a capex-advantaged outlier, with TeraWulf at 5% and Cipher at 4% stabilized ROA—maps to what disciplined underwriting would suggest. In a cycle crowded with copycats, the advantage belongs to those with verifiable cost advantages, bankable contracts, and the patience to price risk rather than chase headlines.