AI Answers in a Second. Your Company's Speed Is Set by Decision Round-Trips: Lessons from a 14,000-Tokens-per-Second Demo

A demo of AI producing text at roughly 14,000 tokens per second has been spreading in short videos. What I want to talk about here is not the chip itself. It is this: once AI becomes this fast, it is no longer AI that sets the speed of a company.

What sets a company's speed is the round-trip of human decisions: waiting for a check, waiting for approval, waiting for the next meeting. If AI answers in one second but deciding whether to use that answer takes three days, the company moves at the speed of the three days. This article continues the October 18 piece, "Compete on the Number of Attempts," and looks at decision waiting.

Read the demo numbers by their source

The clip shows Emad Mostaque, former CEO of Stability AI and now head of Intelligent Internet, trying a chip from the Canadian startup Taalas. Taalas has hard-wired Llama 3.1 8B, a relatively small model released by Meta, into the chip's circuitry. The chip can run only that model, and in exchange it reaches speeds an order of magnitude beyond the usual.

The numbers come from different sources. Taalas's official announcement is 17,000 tokens per second per user (its own measurement). EE Times tried the online demo and observed more than 15,000 tokens per second, and Forbes reported a measured 14,357 tokens per second in its February 19, 2026 article. The same EE Times piece cites figures from Artificial Analysis: Cerebras at close to 2,000 tokens per second and Groq at around 600. Taalas also acknowledges that it quantizes the model aggressively, so speed does not by itself mean answer quality.

It is people, not AI, who are waiting

Suppose a salesperson asks AI to draft a campaign plan. The draft takes seconds. But it then goes to the section manager, the department head and the president, and collecting every sign-off takes a week. If revisions come back, that is another round. AI shortened only a small part of the whole.

Decision waiting falls roughly into four kinds: waiting for the material to be gathered, waiting for the decision-maker's calendar, carrying the item over to the next meeting, and passing it around because nobody is clearly responsible. None of these shrink when AI gets faster. The faster AI becomes, the larger the share of total time taken by waiting on people.

"The longer people deliberate, the better the result" is, I think, a misconception

Some say decisions take time because people are deliberating carefully. In most cases, I think that is a misconception. Truly meaningful deliberation requires experience in the field, broad judgment, technical understanding, tolerance for trial and error, and a medium- to long-term view, all at once. Very few people anywhere have all of these. In my view, most people reach the right conclusion faster by acting along the lines AI proposes.

That does not mean people have no role. People decide what outcome they want, how much risk they will take, and what to try next after seeing the results. I believe human work is shifting from thinking at a desk for a long time to deciding, trying and looking at the results.

Companies slowed by decision waiting lose on the number of attempts

Being slow is not just wasted time. Trying, missing, fixing and trying again is also how people grow; when waiting reduces the number of attempts, people's growth slows too. A company that can try something once a week and one that can try every day end up, after a year, with roughly 50 versus 250 attempts, and the learning compounds. Every company has the same access to fast AI. The difference is how short each company can make its decision round-trips.

In my own company, almost the entire system is integrated. Because decision material is in one place, there is almost no waiting while someone collects numbers. Our marketing now runs on about three minutes of my time. What I do is set the direction, answer questions and check the result at the end; the steps in between no longer wait for my decision each time.

Take risks with a challenger's mindset

On October 15, 2026 (US time), Tesla unveiled a prototype of its new Roadster at SpaceX's test site in Texas. According to reports, it launched off an upside-down ramp, used bursts of compressed air to turn and land, and jumped over a row of vehicles. What it showed were short jumps and flips, not sustained flight. The car had no one on board, production is planned for 2028, the flight-capable version is invitation-only, and it is not approved for public roads (Fortune, CBS and others). It would be wrong to say a flying car has been completed.

Even so, what I notice is the attitude of testing that far and showing it in public. This is my personal view: Japan has many management and work theories that sound right, yet the economy has struggled for a long time and very few world-changing companies have emerged. What matters is not whether something sounds right but the results. I think we should recognize that we are now challengers who have been left behind by the United States, and that this is the time to take risks and increase the number of attempts.

What you can do this week

Pick the decision that was stuck the longest last week and write down how many days it waited. Then choose one decision the president does not need to make, and decide who will own it, how far they may experiment, and how to roll back if it misses. The person given the decision has AI draft options, tries them on a small scale and reports only the results. That alone removes one decision round-trip.

Also count how many new things you tried this month and how many days were lost to decision waiting. Not the number of times AI was used, but whether the number of attempts is rising, is the measure of whether your company is getting faster.

▶ Watch the full video (Japanese)
https://www.youtube.com/watch?v=r_DcHP9lPms