New efficiency breakthroughs could lower AI costs for everyone, but energy strain, grid limits, and war-driven supply disruptions will shape who benefits first.
by Jamie R. Van Doren
What’s Changed?
Last week, Google Research published a compression algorithm called TurboQuant that does something genuinely useful: it shrinks the working memory AI models need to operate — by roughly six times — without degrading performance. On Nvidia’s H100 chips, a version of the technique delivered up to an eightfold speedup in a key processing step. Memory chipmakers noticed. Within hours of the announcement, shares of Samsung, SK Hynix, and Micron all dropped, as traders recalculated just how much physical hardware the AI industry will actually need.
The internet, naturally, compared it to the fictional compression algorithm from HBO’s Silicon Valley. Fair enough. But the business implications are real.
TurboQuant is not an isolated event. It sits inside a broader pattern: AI models are getting smaller, faster, and cheaper to operate. Techniques like quantization, distillation, and compression have been steadily reducing the computing resources needed to run useful AI. What once required racks of specialized hardware is beginning to run on leaner setups: smaller cloud instances, edge devices, even laptops.
Google released TurboQuant under an open research framework, and community developers had already begun porting it to consumer-grade hardware within a day of the announcement. An official open-source release is expected in the second quarter of this year.
Full Credit Google Research
This is the efficiency side of the equation. And for smaller firms, it’s the side worth paying attention to.
But there’s also another side.
Why Does It Matter?
AI infrastructure is under strain. Not eventually. Right now. The U.S. Department of Energy projects AI energy demand will double or triple within the next few years, potentially reaching 12% of total national electricity consumption by 2028. The country’s largest grid operator, PJM Interconnection, which serves over 65 million people across 13 states, has warned it could be six gigawatts short of reliability requirements by 2027. For everyday Americans, that means being told not to run your air conditioner on the hottest day of the year. And if enough people ignore that request, it means rolling blackouts so the whole grid doesn’t go down.
Retail electricity prices have already risen more than 40% since 2019. Some of that is weather, regulation, and fuel costs. But data center demand is an accelerating factor, and utilities from Virginia to Ohio to Texas are scrambling to keep up.
Large tech companies are responding by locking in long-term energy contracts, investing directly in power generation, and competing for grid capacity in ways that smaller firms simply cannot replicate. Meta recently committed up to $27 billion in a single deal for dedicated compute infrastructure. Google, Microsoft, and Amazon are collectively planning hundreds of billions in data center capital expenditure through the end of this year alone.
There’s a less obvious ripple, too. Iranian strikes on Qatari gas infrastructure have knocked out roughly a third of global helium production — a gas most people associate with party balloons but that chipmakers need to manufacture the semiconductors AI runs on. Without helium, you can’t etch the chips that power data centers. Software can get more efficient, but it still needs hardware, and that hardware supply chain just got more fragile. Cheaper algorithms don’t help much if you can’t build the machines to run them on.
So, here’s the tension. AI is getting cheaper to run, yes. But the infrastructure that supports it is getting more expensive and much more constrained. Energy costs are climbing. Grid capacity is tightening. And now, war-driven supply chain disruptions are threatening the materials needed to build the hardware itself. Efficiency improvements like TurboQuant help at the software layer, but software runs on chips, chips run on power, and both of those supply chains just got more complicated. The bottleneck isn’t the algorithm anymore. It’s, well… everything else.
For MBEs, this creates a two-sided opening.
The opportunity: If useful AI tools require less memory and less compute, the cost of adoption drops. You don’t need a massive technology budget to use AI well. You don’t need to build anything from scratch. The tools that help you write faster proposals, run quicker competitive analysis, automate reporting, and tighten forecasting are getting better and cheaper at the same time. That lowers the barrier to entry in a meaningful way.
The risk: Infrastructure advantages compound. Companies that can secure compute capacity, negotiate energy contracts, and invest ahead of demand will operate with structural advantages that have nothing to do with intelligence or effort. If energy costs keep climbing and grid access remains uneven, operational resilience becomes a competitive differentiator, not just a nice-to-have.
It’s worth repeating on point in particular: recent war-driven energy shocks are a reminder that technology growth doesn’t happen in a vacuum. As oil, power, and materials markets tighten, operational resilience matters just as much as innovation. The AI economy depends on the physical economy, and the physical economy is under pressure from several directions at once.
What Should You Do?
For MBEs: focus on disciplined adoption, not ambition.
The advantage right now isn’t in building custom AI. It’s in applying existing tools to real bottlenecks. The companies that benefit won’t necessarily be the ones experimenting casually. They’ll be the ones building repeatable workflows with clear outputs and measurable time savings.
Where are you spending hours on work that a well-configured AI tool could cut by a third? Proposal drafting, compliance documentation, market research, internal reporting? These aren’t glamorous use cases, but they’re the ones that actually change how a small company operates day to day.
The right question isn’t “Are we using AI?” It’s “Can we point to specific outcomes that improved because of it?” If the answer is vague, the implementation isn’t working yet.
For corporate members: start looking at efficiency as a signal.
A supplier using AI well may not look different on a capabilities slide. But they’ll be more responsive. Their documentation will be cleaner. Their turnaround will be faster. Their communication will be more consistent. These are observable differences, and they’re worth weighting in supplier evaluation.
As AI becomes more accessible, the gap won’t be between companies that use it and those that don’t. It will be between companies that use it with discipline and those that treat it as a novelty. That distinction will show up in execution long before it shows up in an RFP/RFQ response.
OMSDC’s Upcoming AI Workshop Series
The Ohio Minority Supplier Development Council (OMSDC) is developing a practical AI series for MBEs and others. We want to hear what formats and topics would actually be useful. Take the short survey at the link below. Complete it and you can download the Competitor Analysis AI Workflow one-pager, a step-by-step guide with prompts and instructions you can put to work immediately.
Take the survey and download the workflow
Questions Worth Asking
For MBEs:
- Where are you spending time that could be reduced by 30–50% with the right tools? And have you actually tested that?
- Are you building repeatable AI workflows, or relying on ad hoc experimentation?
- If energy and compute costs rise, does your operating model stay viable?
For corporate members:
- Are your suppliers becoming measurably more efficient over time, or staying flat?
- Are you evaluating responsiveness and operational clarity, or just price and scale?
- How do you identify partners who are quietly improving their operations through technology?
For both:
- If AI becomes cheaper and more accessible, what differentiates you?
- If infrastructure becomes more constrained, how do you stay flexible?
The Takeaway
The efficiency race in AI is real. And it favors smaller, disciplined adopters more than most people realize. But it’s happening against a backdrop of physical constraints, energy, grid capacity, supply chain friction, that won’t resolve quickly. The companies that navigate both sides of that equation, getting leaner on the software side while staying resilient on the infrastructure side, are the ones best positioned for what comes next. The most advanced AI won’t necessarily be the differentiator. The most disciplined use of it will be.
