Lowering the cost of intelligence forces a shift in how we prioritize asset protection. When the price of compute drops, the barrier to entry for attackers falls alongside it. This accelerates the need to identify the specific data flows that constitute a company hot wallet. If the cost to exploit a vulnerability is negligible, the only defense is knowing exactly which assets hold the keys to the kingdom. We must move beyond perimeter defense toward a model where the value of the data dictates the security posture. This is the natural evolution of the efficiency gains being discussed.
The '13x cheaper, where's the payoff' tension is exactly what CFOs ask. My read is that unit cost fell but usage and review overhead grew, so net spend didn't. Does the report break out any enterprise payback data, or is it mostly lab and vendor numbers?
In consumer goods we see the same thing: a cheaper model doesn't move results if sales and shelf data arrive late or messy. Our biggest return came from automating field-team reporting, not from a better model. Which specific process did your EBIT figure measure?
I felt the 13x from the other end. I tested a new kind of model this month for $1.51, a price that would have been a joke two years ago. What I can't tell yet is whether the cheap test changed what I actually build, which might be the same question your EBIT number asks at company scale.
From a hiring desk, the gap between cheap models and flat profits looks like a data problem. Most of the "AI" roles I filled this year were data roles in disguise: migrations, platforms, Fabric and Power BI. The model got cheap. The plumbing didn't.
The contrast you draw is the part that stayed with me: models getting far cheaper while only a minority of companies see it in profit. That gap makes sense when you separate feeling more productive from actually changing a process. I also found the agent cost point useful — a cheaper single call is not the same as a cheaper finished task when one request quietly becomes dozens. For a smaller team, the practical test you suggest seems right: track the monthly spend and name the process that genuinely changed, rather than judging AI by the price of the model alone.
The 37 percent EBIT figure is the number that matters most here. Cheaper models mean little if only a small share of companies can turn that saving into profit.
The agent cost trap is the part most cheaper-model headlines hide. A lower price per call means nothing if one task quietly triggers dozens of them. Have you seen teams catch that in their bill before it surprises them, or does it usually show up as a mystery invoice first?
The 37% EBIT figure stands out next to where the money is going. PitchBook-NVCA's Q3 2026 Venture Monitor says AI has taken 82.7% of US VC deal value this year, 'the highest annual share in our dataset.' That doesn't prove the payoff gap will close, but it does suggest investors are pricing in gains most adopters can't yet see. Did the report say what the 37% have in common?
The 13x drop holds the score fixed, and plenty of teams don't. They move to the newer reasoning model, spend more tokens per answer, and the bill goes up anyway. Tracking cost per resolved task month over month shows whether the falling price ever reached your P&L.
That 37 percent EBIT figure barely moving in a year, while lab revenue runs from about $30B to $105B, hints at where the gains are landing so far. Cheaper models help, though most companies still seem stuck on integration, data quality and governance, the unglamorous parts. Only three labs left at the frontier deserves more attention too. Concentration like that can leave the American lead exposed to a single bad outage, lawsuit or security breach at one firm, and Beijing is surely watching for exactly that kind of opening. Benaich grading his own past forecasts in public is a good habit, and one more analysts could pick up.
Lowering the cost of intelligence forces a shift in how we prioritize asset protection. When the price of compute drops, the barrier to entry for attackers falls alongside it. This accelerates the need to identify the specific data flows that constitute a company hot wallet. If the cost to exploit a vulnerability is negligible, the only defense is knowing exactly which assets hold the keys to the kingdom. We must move beyond perimeter defense toward a model where the value of the data dictates the security posture. This is the natural evolution of the efficiency gains being discussed.
https://cyrilsimonnet.substack.com/p/every-company-has-a-hot-wallet?utm_source=substor&utm_medium=substack&utm_campaign=comment
The '13x cheaper, where's the payoff' tension is exactly what CFOs ask. My read is that unit cost fell but usage and review overhead grew, so net spend didn't. Does the report break out any enterprise payback data, or is it mostly lab and vendor numbers?
In consumer goods we see the same thing: a cheaper model doesn't move results if sales and shelf data arrive late or messy. Our biggest return came from automating field-team reporting, not from a better model. Which specific process did your EBIT figure measure?
I felt the 13x from the other end. I tested a new kind of model this month for $1.51, a price that would have been a joke two years ago. What I can't tell yet is whether the cheap test changed what I actually build, which might be the same question your EBIT number asks at company scale.
From a hiring desk, the gap between cheap models and flat profits looks like a data problem. Most of the "AI" roles I filled this year were data roles in disguise: migrations, platforms, Fabric and Power BI. The model got cheap. The plumbing didn't.
The contrast you draw is the part that stayed with me: models getting far cheaper while only a minority of companies see it in profit. That gap makes sense when you separate feeling more productive from actually changing a process. I also found the agent cost point useful — a cheaper single call is not the same as a cheaper finished task when one request quietly becomes dozens. For a smaller team, the practical test you suggest seems right: track the monthly spend and name the process that genuinely changed, rather than judging AI by the price of the model alone.
The 37 percent EBIT figure is the number that matters most here. Cheaper models mean little if only a small share of companies can turn that saving into profit.
The agent cost trap is the part most cheaper-model headlines hide. A lower price per call means nothing if one task quietly triggers dozens of them. Have you seen teams catch that in their bill before it surprises them, or does it usually show up as a mystery invoice first?
The 37% EBIT figure stands out next to where the money is going. PitchBook-NVCA's Q3 2026 Venture Monitor says AI has taken 82.7% of US VC deal value this year, 'the highest annual share in our dataset.' That doesn't prove the payoff gap will close, but it does suggest investors are pricing in gains most adopters can't yet see. Did the report say what the 37% have in common?
The 13x drop holds the score fixed, and plenty of teams don't. They move to the newer reasoning model, spend more tokens per answer, and the bill goes up anyway. Tracking cost per resolved task month over month shows whether the falling price ever reached your P&L.
That 37 percent EBIT figure barely moving in a year, while lab revenue runs from about $30B to $105B, hints at where the gains are landing so far. Cheaper models help, though most companies still seem stuck on integration, data quality and governance, the unglamorous parts. Only three labs left at the frontier deserves more attention too. Concentration like that can leave the American lead exposed to a single bad outage, lawsuit or security breach at one firm, and Beijing is surely watching for exactly that kind of opening. Benaich grading his own past forecasts in public is a good habit, and one more analysts could pick up.