Last week I argued that most PE portfolios aren't getting AI operating leverage because they treat AI as a tool when they need to treat it as an operating system. That gap already has a price in the valuation multiple.
This week, Forbes gave it a second price: the energy bill.
In Why AI Efficiency Is Now A Board Governance Issue (Forbes, Sep 21, 2026), board director Heather Wishart-Smith cites the International Energy Agency: electricity used by AI data centers rose roughly 50% in 2025 alone. Lawrence Berkeley National Laboratory projects U.S. data centers could consume 6.7% to 12% of all U.S. electricity by 2028, up from about 4.4% a few years earlier.
Her conclusion is the right one. AI efficiency has moved off the IT budget and onto the board agenda. A board that can't explain, layer by layer, where its AI footprint comes from isn't governing it.
I'd go a step further. The biggest AI efficiency problem isn't compute. It's context. And the fix is the same Operating Foundation I wrote about last week.
The Forbes piece splits AI efficiency into three layers. Each has its own expert and its own headline number.
Look at where those gains came from. Two of the three had nothing to do with better chips or a better model. They came from sending less context, and better context, into every AI call.
That's a governance problem wearing an engineering costume. You can't send only the relevant context if nobody has defined what's relevant, verified that it's current, or decided who owns it.
Hoang's sharpest line in the article isn't about tokens. It's this: "What's better than reducing tokens? It's not using tokens at all."
He's talking about AI workslop: polished-looking output that turns out to be vague, irrelevant, or wrong, because nobody defined the problem before prompting.
Last week I described a CRO asking her AI tools why deals were slowing down and getting confident, plausible, and completely wrong answers. The tool wasn't broken. It was ungrounded. It had no verified record of her ICP, her qualification criteria, or her pipeline, so it filled the gap with something that sounded like an answer.
That's workslop. And workslop costs you twice:
Then the team re-prompts, attaches more documents, and drags more history into the next call. The token bill grows and the answer still isn't grounded.
The guardrail I described last week is the most efficient thing in the stack: an AI skill checks its dependencies before it runs. If the qualification criteria, positioning, or pipeline data it needs hasn't been verified as current, the skill doesn't fire. You get no output, just a flag saying what needs to be verified first.
That's Hoang's point turned into a mechanism. The cheapest token is the one a governed system refuses to spend.
The three components from last week, Foundation, Layer, and Analytics, line up closely with what Forbes says boards now need to govern.
The architecture that stops AI from hallucinating is the same one that stops it from wasting tokens. They're not separate initiatives. They're one discipline.
This is where it matters most for private equity.
Last week I described a fund where every portfolio company builds and rebuilds the same operating foundation (qualification criteria, lifecycle definitions, playbook logic) on its own. I called that a missed synergy. It's also a duplicated energy bill.
Every portfolio company without a governed Foundation is:
Gale's warning applies here: an unoptimized query doesn't run once, it runs millions of times a day as agent activity grows. Multiply that by every company in the fund.
A portfolio-wide Context Vault flips the math. The fund defines and verifies the Foundation once, and every company inherits it. Skills run on context that's already scoped and current. The efficiency gains compound across the portfolio the same way the operating gains do.
When I ran portfolio companies for OMERS Private Equity, the discipline that kept strategy honest wasn't a policy document. It was a cadence: Hoshin Kanri, with clear objectives, owners, measures, and regular reviews where assumptions got challenged. AI efficiency needs the same treatment.
Forbes suggests boards ask vendors, "What am I actually getting per watt, not per dollar?" Good question. Here are five more for the operating side:
A board that gets clear answers is governing AI. A board that gets blank looks is funding experiments.
Last week's post ended with a binary choice: build a portfolio-wide AI operating system, or keep funding independent AI pilots.
This week adds a second reason the choice matters. The same Foundation that turns AI Users into AI Operators, and that McKinsey ties to a roughly 130% higher median revenue multiple, also cuts the tokens, power, and rework behind every AI interaction.
Efficiency and operating leverage aren't competing priorities. They come from the same three components working together:
Scrutiny is coming, from regulators, LPs, and buyers at exit. The funds that can explain their AI footprint layer by layer will have an answer. The rest will have an energy bill and a slide deck.
Ask your board which one they have.
Let's Connect → See how a portfolio-wide Context Vault makes AI both more effective and less wasteful.
