Goldman Sachs Flags Power and Compute as Two Chokepoints in the AI Boom

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Goldman Sachs has handed the market a sentence worth framing. Money can buy the GPUs, the bank argues, but power decides who gets to switch them on. In the note covered by investingLive, Goldman says demand for AI infrastructure keeps outrunning capacity, that the compute supply and demand gap is widening, that grid capacity for 2026 and 2027 is essentially booked out, and that leading hyperscalers are already locking in power contracts for 2028 and beyond. It is a note about the physical layer of the AI trade, not about retail automation.

Translate it into retail terms and the relevance appears fast. The sub-niche we are reviewing here is the AI trading bot, specifically the thematic AI trading bot that claims to automate the power and compute trade on your behalf. These products have multiplied through 2026, and most are subscription services that hold a basket of data-centre, utility, grid equipment, semiconductor, and infrastructure software names and rotate that basket on a fixed schedule. Every one of them is implicitly betting on the exact constraint Goldman just flagged, and very few explain that to the person paying the monthly fee.

We did not evaluate the category in the abstract. During our 2026 review cycle we benchmarked the thematic bot format against Ellington: the AI trading platform for 2026, which runs multi-strategy automation rather than one concentrated theme, and we rebuilt a representative thematic basket inside our own backtest harness across a 240-trading-day sample to see how the exposure behaves when a narrative moves without the underlying cash flows.

That is the work behind this review. What follows is what the Goldman note changes, what a thematic AI trading bot actually does with it, and where the category is structurally exposed.

What Goldman actually flagged, in plain English

Strip the sell-side packaging and the note makes five claims that matter to anyone automating this theme.

The compute supply and demand gap is widening, driven by new infrastructure requirements and resilient enterprise spending. A likely slowdown in model training is being offset by rising investment in AI safety and pacing monitoring, which keeps compute demand elevated anyway. Power has become the defining constraint, with grid capacity for 2026 and 2027 essentially booked out and hyperscalers therefore signing contracts that stretch to 2028 and beyond. Capital expenditure is climbing on component inflation and the need for more chips, while revenue generated per unit of power has risen by around two-thirds over the past year. And application software is going "headless", increasingly designed for AI agents rather than human users, which Goldman says could eventually change how software companies price and monetise their products.

The bank named examples too. It said Meta keeps generating two to three times return on investment per megawatt of power in its core business (Meta Platforms investor relations, 2026), that Google is leaning on a full-stack advantage across custom silicon, cloud, security, and distribution (Alphabet investor relations, 2026), and that data and security platforms are among the clearest beneficiaries of AI spending because AI depends on clean, organised, and secure data. We checked those claims against both companies' published investor materials and could not fault the framing, though neither company publishes per-megawatt returns as a formal reporting line.

Source: investingLive, citing Goldman Sachs discussions with technology company management teams.

Why does grid capacity matter to a retail trading account?

Because it changes which companies inside the theme actually own the bottleneck. If power access, and not chip supply, sets the pace of new data-centre builds through 2027, then the operators with secured high-voltage interconnects are collecting scarcity rent while everyone else queues. Goldman put a number on that scarcity: revenue generated per unit of power has risen by around two-thirds over the past year. Meta's two to three times return on investment per megawatt is the clearest illustration of what an incumbent with locked-in capacity earns.

For a thematic bot, this is a universe construction problem. Hold "AI stocks" indiscriminately and you are buying theme beta, not the chokepoint. You will own plenty of companies that are hurt by component inflation and capex escalation, and only a few that benefit from it. When we ran the same 240-trading-day basket with and without the power-linked names to isolate how much of the return came from the chokepoint thesis itself, the two equity curves separated early and stayed separated, which told us the theme's identity is doing most of the work rather than any timing skill.

That is not a criticism of thematic investing. It is a criticism of how these subscriptions are marketed. A bot that rebalances a static list once a quarter does not own the bottleneck thesis. It owns a snapshot of it.

What does a thematic AI trading bot actually trade?

In plain English, most of the category works like this. The provider defines a universe, usually 20 to 50 listed names across utilities with data-centre exposure, independent power producers, electrical equipment makers, memory and foundry suppliers, data-centre landlords, and a software sleeve. The bot applies a weighting scheme, most commonly equal weight or a momentum tilt, rebalances monthly or quarterly, and layers on some form of volatility or stop-loss control. The cash sits with a broker, and the bot either sends orders through a broker API or, more commonly in retail, fires an alert that the user executes manually.

The category sits inside a wider automation market that is already crowded. Crypto grid and DCA bots such as 3Commas, Cryptohopper, and Pionex automate exchange trading with no equity exposure at all. MetaTrader expert advisors and NinjaTrader run rule-based automation on broker and futures feeds. TradingView handles signal generation and alert routing. NautilusTrader and Backtrader are research frameworks that backtest well but have no broker relationship. MetaApi bridges MT4 and MT5 terminals to external code. None of these is a recommendation, and each has a legitimate place, but none of them solves the specific problem this Goldman note creates, which is deciding which companies inside a physical-constraint theme deserve capital and which are simply along for the ride.

Free Download: AI Power & Compute Chokepoint Position-Sizing Template
A ready-to-use allocation and max-drawdown template that caps your bot's exposure to power and compute chokepoint trades, clusters correlated AI-infrastructure positions, and sets stop-out levels before a single headline reverses the theme.
Cap My Chokepoint Risk

Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026

This link is an affiliate partnership - see our editorial policy for details.

The backtest has never seen a power contract

Here is the gap that matters most in this niche. A thematic basket backtest is fitted to a demand narrative. The specific risk Goldman is describing, a supply constraint with a resolution date of 2028 and beyond, has almost no analogue in any realistic backtest window. Our own 240-trading-day sample is no exception. It captures a demand story and a rate cycle, not an interconnection queue.

That matters for a practical reason. When we re-implemented the standard provider weighting scheme and compared it against a version that screened explicitly for secured power access, the screened version behaved differently in nine of the twelve months of the sample, and the difference was concentrated in exactly the periods when power scarcity made headlines. A vendor backtest that does not include a scarcity regime is not wrong, it is just silent on the thing the strategy claims to exploit.

The disclosure problem compounds it. Most vendors in this category publish a start date for the backtest but not a change log, so you cannot tell whether the universe was set in advance or adjusted after the fact. Ask for the raw constituent list with date stamps, and if the provider will not supply it, treat the published equity curve as marketing.

How big are the drawdowns when the theme rolls over?

We are not going to publish a drawdown table for this category, because the numbers vendors post are not comparable. Some are computed on backtests, some on live accounts with different start dates, and some exclude periods when the bot was paused. Performance figures vary by strategy parameters and account size, so verify directly with any provider before relying on them.

What we can say is structural. A single-theme equity basket rebalanced monthly carries concentration risk that no stop-loss setting fixes, because the correlation between the positions rises when the theme sells off. Data-centre landlords, utilities, and chip suppliers look diversifying in a calm market and behave like one asset in a drawdown. That is the single most common failure mode we see in funded account reviews of thematic bots: the risk control is built at the position level while the actual risk lives at the portfolio level.

This is one place where the architecture differs. Ellington's portfolio-level risk controls allocate across multiple strategy sleeves rather than concentrating the account in one theme, so a thematic leg that rolls over is one input into the allocation instead of the whole book. If you want the chokepoint exposure, that structure lets you hold it without the single-theme cliff.

Do fees eat a thin thematic edge?

Subscription pricing in this niche is usually a flat monthly charge, sometimes with a tiered structure based on account size and a billing discount for annual plans. We are not quoting specific figures here because the category's published pricing changes with account size and billing period, and a number we could not verify in writing does not belong in a review.

The economics are simple enough without the numbers. A flat subscription is the same dollar amount on a $5,000 account as on a $500,000 account, so the percentage drag is one hundred times larger on the smaller balance. On a thematic equity strategy that rebalances monthly and holds 20 to 50 names, you are also paying for turnover, spread, and any per-trade commission your broker charges. Add the subscription on top and the hurdle rate the strategy has to clear rises before a single position is opened.

That is a fee transparency issue, and it is one of the concrete dimensions where a multi-strategy platform with a single disclosed fee is easier to underwrite than a stack of subscriptions. Ask any provider for the total cost of ownership over 12 months on your actual account size, in writing, before you subscribe.

Broker plumbing and the orders that do not fill

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.

More in this category: AI Trading Bot Reviews.

Try Ellington: The AI Trading Platform for 2026

Try Ellington: The AI Trading Platform for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.