Everyone Says Slow Down, Nobody Stops: AI Safety as a Game, and a Trade Show Full of Undecided Technology
Notes after listening to the Statementdog podcast (2026-09-17): why the market shrugged when AI labs collectively called for slowing down, what compliance costs actually looked like in banking and crypto, and what an optical trade show with no settled standards means for reading the industry. Educational, not investment advice.
Contents
A man of Chu was selling shields and spears. He praised his shields: “My shields are so solid that nothing can pierce them.” Then he praised his spears: “My spears are so sharp that nothing can withstand them.” Someone asked: “What if your spear were driven against your shield?” The man had no answer.
—— Han Feizi, “Nan Yi” (Warring States period; translation mine)
The man from Chu had nothing to say, because both sentences were his own. Listening to the Statementdog podcast this week (2026-09-17, episode 565) on AI safety and an optical trade show, that scene is what came to mind: a group of people out in front saying “we should slow down” and “we intend to win” in the same breath, with the market watching and believing neither.
What the episode covers
The first half starts with Anthropic’s CEO calling for a slower pace of AI development, followed by OpenAI, Google DeepMind and Musk echoing him — close to every head of a leading American model lab endorsed the idea. Markets got nervous for a moment, since this boom has been holding up equities for years. The host asked Sky, the investment director, what he made of it. His answer: it smells like a fake issue.
The second half moves to CIOE, the China International Optoelectronic Exposition. Sky did not attend; friends who did brought back the observation that the venue runs about nine times the combined floor area of Taipei Nangang’s two halls, with a single theme swallowing an entire building. What people were arguing about there ranges from CPO versus NPO, to whether the light source should be DFB or VCSEL, down to which test stations survive.
The points worth keeping
1. The call to slow down is a move in a game. Sky put it plainly: the runner in first place turning around to tell everyone to slow down is funny on its face. Even if the leaders would like to slow — training burns money, inference earns it, and a collective slowdown would let everyone book profit — who dares stop first? If one rival slows less, the position changes hands. And the table only holds American firms. China is not stopping, and users will not switch to a model because it behaves better. Sprinting is each player’s best response and also the equilibrium.
2. The signal from the Trump administration is “no regulation.” The episode mentions Trump phoning Jensen Huang, who was on stage, on speaker. The content of the call matters less than the atmosphere it happened in: political appetite for regulation and moral appeals from vendors are separate things.
3. Compliance cost has existing reference numbers. Finance is the heavily regulated industry — examinations arrive constantly, and even a fund buying stocks files buy and sell reports. The magnitudes Sky pulled out: compliance runs around 1% of revenue at large banks, with some estimates at 3%, and up to 5% where examinations are frequent. The emerging-industry comparison is crypto, which went from nobody’s jurisdiction to governments arriving; Binance has disclosed something like three in ten staff working on regulatory requirements. Map that onto AI and the conclusion is that regulation raises cost, and firms with high revenue and high per-token pricing can absorb it. Trouble for the back of the pack, moat for the front.
4. The market barely moved, because no rule text exists. Regulate what, safe by whose standard, how safe is safe enough — nobody has a concrete procedure. Sky’s description of the sentiment: too many Terminator movies. Monday showed a flicker, the next two days nothing. He thinks rate expectations may matter more; rumors like Intel hitting 80%-plus yield on 2nm arrive daily and belong to sentiment trading, so hold onto fundamentals.
5. The uploaded fruit fly. The strangest segment: Google published the most complete fly brain connectome model to date, and the internet promptly wired it into games — Doom, Minecraft, and someone had it mine Bitcoin. Sky finds this more science-fictional than Skynet, though he reads the direction as different: that model develops inside an existing biological wiring diagram, while today’s language models took another road. As for whether simulating a human brain would beat current models, his answer was that the person being simulated might turn out dull.
6. Nothing at the trade show is settled. CPO and NPO substitute for each other, and whether NPO sits on substrate or PCB is open. DFB leads among light sources, some push VCSEL, some want micro LED, and the three put their costs in different places. Whether CPO reaches passive alignment is the hinge — active alignment burns time, which means alignment machines, which sustains an entire equipment tier. Even the test flow keeps moving: originally four stations, then someone added a fifth for burn-in, and this show brought word that station two gets dropped with only three and four remaining. Sky’s conclusion on all of it: he has no conclusion, because the field is still expanding.
7. One chain you can trace. The light source choice rewrites the shortage story: DFB consumes indium phosphide, VCSEL consumes gallium arsenide, so a shift toward VCSEL loosens indium phosphide. The real volume trigger for these components is 1.6T applications arriving — this year’s volume fell short of estimates, and hope moved to next year. Thermal management is changing level too, from system to chip. On the power side there is a phased path toward 800VDC, with silicon carbide migrating from automotive high voltage into the data center. Elsewhere, Infineon sold its NOR and NAND business to Winbond, which Sky reads as capacity being redivided: a company whose core business is overloaded cuts what is not core.
Going further
”The news says AI is cooling. Should I get out?”
That is the reflex after the first segment. It crossed my mind too, until Sky’s line — who dares slow down on training — pulled me back.
In a prisoner’s dilemma, talk is free and action costs. So read declarations and behavior separately: a declaration tells you how someone wants others to act, behavior tells you how they plan to act. A group of competitors calling for a slowdown at the same time tells you they all know their rivals will not slow.
What to watch instead: things that cost something. Capex numbers, data center groundbreakings and leases, wafer and optical module orders, hiring in the critical roles. A slowdown begins existing when those change. My own mistake came after one “the industry is decelerating” report, when I cut half a position and then watched capex guidance revised upward for two straight quarters. That report contained no false statement. My error was using a mood where a number belonged.
The failure condition deserves writing down as well: once slowing down turns from appeal into statute, language acquires a price and behavior follows. So the thing to track is the appearance of penalties, timelines and defined scope, rather than who said what.
”Regulation sounds bearish, so why are the leaders relaxed?”
Instinct says tighter rules mean thinner profit for all. The numbers Sky cited point elsewhere.
Compliance spending behaves like a fixed cost — a legal team, an audit process, a filing system — with weak dependence on revenue scale. A firm with ten billion in revenue absorbing 3% and a firm with one billion absorbing the same machinery feel it an order of magnitude apart. The Binance disclosure stings because those three in ten people produce no new product.
So the day rules land, the market reshuffles relative order. Three questions I would ask: whether this company’s per-unit pricing supports the extra cost, whether current margins leave room to swallow it, and whether its customers become harder to poach once a compliance bar exists. The third gets skipped, and it is the hinge where cost converts into moat.
The opposite case exists too. If regulation takes the form of capping compute or capping model capability, the leader loses the one advantage it has, and the news turns bearish for it. Whether regulation helps the leader depends on whether it governs process or ceiling — process raises the barrier, ceilings flatten the gap.
“So many acronyms. Which one do I follow?”
CPO, NPO, DFB, VCSEL, passive alignment, MSAP, 800VDC — a dozen terms in one episode, and what survives is the feeling of being lost. My approach is to put them all under one question: has a standard been settled?
Sky’s line — there are so many solutions because nothing has been decided — is the key to the segment. While the standard is open, every route gets demonstrated, every vendor claims theirs is best, and the supply of stories exceeds what anyone can read. A supply of stories and a supply of revenue are different quantities. After the decision, the winning route ramps and the rest go to zero.
Standing at an undecided moment, I put attention on three verifiable things. First, whether a volume date has been stated concretely — the wording here on 1.6T is that next year looks like it may step up, and last year’s version of that sentence about this year did not hold. Second, who the customer is and what share of revenue they represent, since under one upgrade cycle the benefit depends on customer and volume rather than on the technology. Third, whether a mechanical checkpoint exists — the test station question is one, because dropping station two rewrites content value for test equipment vendors, and that answer shows up at the next show or earnings call, where you can check your work.
The parts the speaker himself calls undecided go on a watch list, not into a thesis. In this episode that covers CPO versus NPO, the three-way light source contest, and how chip-level thermal management gets done. Admitting something has no answer costs less than inventing one.
Where to look next
- Statementdog podcast episode 565, 2026-09-17 (the source for this piece, available on major podcast platforms)
- The CIOE official site — the theme layout and exhibitor list are primary material for seeing how the technology routes are distributed
- Optical networking standards bodies and their white papers (OIF, OCP and others) — the phased paths for CPO, NPO and 800VDC are public documents
- The FlyWire connectome project’s public data, if you want to see the fly brain that ended up playing video games
- Annual examination reports from financial regulators, plus compliance disclosures from crypto exchanges — the real ratios of compliance cost to revenue live there
The one thing to take away
A public declaration tells you how the speaker wants others to behave. To learn how they plan to behave, look at what they paid.
The episode demonstrates this from end to end. A group of labs said slow down, and none of them cut training spend. A trade show filled with vendors claiming their route is best, and no route has been chosen. Words are free, so words are cheap. Committing money, headcount, long-term contracts, or cutting a product line — those hurt, so those carry information.
Here is something I tried that you can run this week. Pick a statement someone near you made in public recently — a family member saying they will quit smoking, a colleague saying they will deliver next week, a friend saying they will change jobs next year, a boss saying the company will pivot. Write one line in your phone: “If this is true, what specific action shows up in the next seven days?” Specific enough to tick off. “He’ll be more motivated” does not count. “He threw the pack away” does. Seven days later, open that line and check.
The exercise surfaces two things. One, plenty of statements produce a reaction in you while yielding no matching action, which means they carried no information to begin with. Two, once in a while you write it correctly and know the outcome before the person says it. Both get used far more often outside investing than inside it.
This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.