The Question That Keeps Ambushing Us

As we prepare to launch a second fund, a set of questions keeps ambushing us. What do we know that is not known to others, and from this, what does the world know, or what do we know that the world knows? We can go on, but we are assuming by now our readers get the drift.

Our current fund rests on the belief that innovation is one of the few genuine high-growth themes left in a world where globalization and demographics have lost some of their force. It invests in the new-era innovators that turn invention into income. Genuine innovation, but also genuine money-making.

The new fund, GenInnov Epicentre, to launch on September 1, begins one step later. Those who make most money, spend most money. More precisely, those whose incomes grow fastest often have costs that grow fastest too. They build data centers and factories. They consume materials, electricity, logistics and services. Their employees acquire incomes and wealth that alter the neighborhoods around them. The companies constructing those projects, supplying that operating expenditure, or serving those employees can grow in the economic shadow of innovation without being innovators themselves.

This is why Taiwan or Korea can feel as alive today as Shanghai or Bengaluru once did. It is also why, while examining hundreds of companies and dozens of unfamiliar subsectors, we keep encountering things that are new to us.

And then we keep asking, so what? What do we know or can we know of any relevance that is generally not known to others? This essay is an inquiry into knowing. It is an epistemic patchwork rather than an argument, and it settles nothing. But when there is more to know than anyone can know, working out what deserves the attention is not a digression from investing. It is most of it.

Where Serendipity is The Product

The search for something new to us and new to all has been so easy with the innovation fund. 

Innovation has always run on serendipity, and much more so where the underlying sciences are heuristical rather than deductive. Quantum physics had it. Model-making has it now. Nobody builds a frontier model knowing what it will be able to do. They build it, they train it, and then they find out, at roughly the same moment as everybody else. Serendipity is not an embarrassment to heuristical science. It is often the method leaving room for reality to answer back.

Material new information arrives constantly as we monitor the themes of interest. Not everything may be headline-grabbing, but one is constantly processing relevant information that is as new to the people producing it. Take HBM4E. That SK Hynix will spend enormously to get there is known. You could have written that sentence a year ago and been no less right. What is not known, by us, by the analysts, by the customer, and even by the engineers, is when it becomes viable, at what yield, at what price, and through whose qualification first. Expectations exist. Yardsticks exist. The answers do not exist until they do for all.

Investing, as a result, is not second-guessing what one sees as new knowledge. 

Monopoly Power and the Morning Newspaper

In the non-innovation segments our new fund now examines, the central problem is materiality. Genuine newness need not come from products or research; it can arrive as a new financial number, a new business, a new relationship, new margins, new costs. 

But some information has no real newness at all. Articles now regularly marvel that certain companies upstream of technology hold monopoly power over key materials. The chokehold is usually decades old, and its newness might be only due to the discovery by the writer penning it. But then, the newness could be in the company’s newfound capabilities in using the pricing power. 

When we hear of possible peptide shortages driven by changing drug administration methods in attempts to side-step the patent cliff, or new oral GLP-1s, the information may look new to someone like us seeing it for the first time. We have to recognize that the phenomenon has likely been known inside the peptide industry for a long while. The same headlines have circulated before. The same analysts have fretted before when we were busy elsewhere. The same capacity constraints have been modeled, debated, and priced in by the specialists who live in that world.

Even so, if a genuine capacity bottleneck results that cannot be eliminated for years, its consequences are not only prolonged, but in many ways cannot be fully known or priced for an extended period. 

Which leaves three questions where most investors ask one: is this new to me, is it new to the world, and is it new to the price? They answer differently with distressing regularity, because the market persistently confuses the age of a fact with its economic relevance. A structural advantage can be thoroughly documented, entirely known, and carry no premium whatsoever, for no better reason than that the power of the information had no evidence to back it until everything suddenly changes.

Acting on it requires an awkward posture. You must accept that you have discovered nothing original while holding, firmly, that everybody else's indifference is a mistake. There is no glory in it. The return comes not from knowing something first but from being willing to hold a stale fact for the time it takes the crowd to find it interesting. To claim that stale information is mispriced may sound like overconfidence, bordering on arrogance, but in the non-innovation field, that is perhaps where the fund manager is adding most value. 

The Average Is the Least Interesting Number

Everything in markets must be against what is known to the market. And what is known to the markets gets measured through consensus, and what consensus hands us is a single average, expressed to one decimal place, with all the moral authority of arithmetic.

The dispersion around it, and the speed at which it moves, tell you far more. Last year we were repeatedly baffled by high-growth Taiwanese companies whose forecasts sat, year after year, at a dutiful ten to twenty per cent. Today, with the market higher and the growth cemented, the same names carry forty and fifty per cent trajectories. Some of that is analysis, but a lot is likely extrapolation wearing analysis as a coat.

Korea makes it plainer. For several of the memory names, a target price is largely a function of the date on the note: a thirty or forty per cent premium to the prevailing price if in the preceding periods the market trend was up and almost “sell”-appearing 15-20% if market trends have been down. 

Having once run research, I can offer a small confession on the industry's behalf. Where internal rules force a target-price review with every publication, the most prolific analysts are structurally disadvantaged. Write often enough about a fast-moving stock and your numbers begin to follow the price rather than lead it. Nobody intends it. But a strict team head may not approve notes significantly away from the current price.

And this is not confined to target prices. The dispersion in year-two and year-three forecasts, precisely as forward multiples became the yardstick of choice, makes a quiet mockery of the phrase "the market knows".

Who Whispered?

Worse than consensus is the market's newest proxy for knowledge. Ahead of any significant result, a second number now circulates, and beating the official one has stopped being sufficient. A company must also clear a figure that appeared from nowhere a few days earlier, carries no source, and belongs to nobody.

Consider what this means in an industry where compliance will not let an analyst publish a bar chart without a citation beneath it, down to "House Estimates" for numbers the house invented itself. Attribution is enforced everywhere, on everything, by people with the authority to stop the presses. And then, once a quarter, a number with no author, no methodology and no owner is allowed to set the bar for a company worth fifty billion dollars. The whisper is the only figure in finance exempt from the rules that govern every other figure in finance.

We have tried to establish where they come from and have failed. Our best reconstruction runs as follows. A salesperson of sufficient standing floats a figure. Three or four people are consulted. Each glances at consensus, adds a markup calibrated to the mood of that morning, and hands it back. Someone writes it down. A journalist repeats it. By Thursday it is the expectation, and by Friday somebody has lost money against it.

The technique has since escaped earnings season. Anyone wishing to assert something about an innovation timeline, without the inconvenience of evidence, need only phrase it as what the market is quietly expecting. Whispering is the most efficient form of authority yet devised: no name attached, no proof required, and nobody available to be wrong.

When No One Loses

Technology is an industry that talks almost exclusively about tomorrow. Every quarter arrives with new products, new versions, new acronyms, and a newcomer's certainty that something unprecedented is underway, when in several of these segments the same “unprecedented” progress as well as adjective usage has been underway for thirty years.

The more interesting distortion is not the jargon. It is that "what everyone knows" changes depending on whose desk you are standing at.

Optical interconnect is the cleanest current example. Several rival approaches are alive at once, and most of them must die for one of them to matter: pluggables, linear and near-packaged optics, co-packaged optics, and the stubborn survival of copper over short reaches. Speak to the analyst covering the copper beneficiaries and co-packaged optics is definitively delayed, with cost and thermal problems nobody has solved. Speak to the analyst covering the co-packaged names and adoption is running ahead of plan and the addressable market is being revised upwards.

Both will present the position that suits the conclusions they have predecided as consensus. Often they work at the same firm, if not in the same team. Occasionally they are the same person, on different days, covering different stocks. What is knowable, it turns out, has a seating plan.

Three Years Behind What?

Late converts make the loudest zealots. Analysts who dismissed Chinese innovation outright now credit CXMT with capabilities they withheld for years from manufacturers carrying decades more scar tissue. Lithography will receive the same treatment shortly. Because Chinese model-makers have closed on OpenAI and Anthropic faster than almost anyone expected, it has become fashionable to assume the same compression applies to ASML or Hynix, as though a transformer and a lithography scanner were the same species of problem.

Whether Chinese memory and foundry manufacturing is closing on the frontier is a serious question, and it deserves genuine re-examination every few quarters. The unseriousness begins with the answer. "Three years behind" is the standard figure, and it drifts, on very little new information, according to the temperature of the week. Behind on which process, at what yield, on whose equipment, qualified by which customer, and measured on what date? A single number answering all five is not an estimate. It is a mood with a decimal point.

China's investment and its progress are both real. What is underestimated is the speed of the thing being chased. Roughly a hundred billion dollars over more than a decade brought its industry to where it stands, which is a genuine achievement. But the frontier is not waiting on a platform for China to arrive. It is moving, at present faster than at any point in the industry's history, funded by profits and by a confidence in research returns that did not exist three years ago. The gap may well be widening rather than closing.

We are fond of comparisons built to shock, because large numbers otherwise induce a comfortable numbness. We floated the idea that Hynix's profits might exceed Nvidia's when consensus forecasts sat at a fraction of it, and early this year we suggested that Samsung's 2026 profits could exceed the combined profits of all listed companies in India. Here is the next one. A single memory maker's projected capital expenditure over five years may exceed everything China has invested in semiconductors to date.

None of these comparisons predicts a winner. They are arguments against confidence in either direction.

Vulnerability of the Settled

Producing high-quality research on dozens of companies a day has become close to trivial. Our own stack now generates, in a few weeks, more genuine analytical material than what a large sellside house could create. With so much internal material, over and above the rising volume from everyone around, one has to revive an old question in a new setting: if the research is written and nobody reads it, has anything been added to what the market knows?

Consider how much fully public material is effectively unread. The two-hundred-page analyst day deck, downloaded widely and finished by perhaps a dozen people. The sixty-page initiation note whose executive summary is quoted for years while pages nine to fifty-one remain pristine. The celebrated long essay, forwarded within an hour of publication, which is roughly the time required to read its first three paragraphs. Availability has been mistaken for knowledge for as long as documents have existed, and the mistake has worsened now that forwarding is free while reading still costs an afternoon.

Then the harder case, where the document was genuinely read. "Attention Is All You Need" was published in 2017, free and public, and everyone in the field read it. Markets ignored it for five years. ChatGPT arrived, and the same eight pages became the most consequential document of the era. Nothing in the paper had changed. What arrived was the evidence that an obscure set of equations worked, and with it the probability the world had declined to assign.

The traffic now runs both ways. Nature reported this month that AI agents auditing the scientific literature are surfacing errors that sat undisturbed for decades. A chemist at Zhejiang Lab found his model contradicting a seventy-five-year-old reference database on boiling points, assumed the model was wrong, checked the original literature, and found that the database was. Agents rerunning the claims of papers accepted for oral presentation at a leading machine-learning conference could reproduce the great majority of headline results in only a handful of cases.

We see a smaller version daily. Our agents read what they are permitted to read and return, unasked, with a large list of inconsistencies in published material every day. We often read famous commentators waxing eloquent about LLM bias and hallucinations even now to conclude how they would never rely on machines’ research for any serious work. The evidence of who makes more errors and whose errors could be easily spotted points more and more in the other direction. 

Which leaves the investor short of a respectable excuse. What we file as consequential is rarely a verdict on the information. It reflects our priors, the mood of the week, who sent it, how the first line was phrased, and now, increasingly, which model we happened to ask or which agent was scheduled to monitor.

When the Fragments Speak

Outside the laboratory, information matures differently. In hardware supply chains, an announcement marks the moment a fact becomes official. It rarely marks the moment the fact became inferable.

The production schedule of a dominant chip designer is not a sealed secret until an executive reads a script on an earnings call. It exists weeks or months earlier, in pieces, distributed across test and packaging houses, substrate makers, memory suppliers and the firms selling liquid cooling. A simultaneous acceleration in high-bandwidth memory orders and cooling components does not hint at the customer's revenue line. It very nearly states it. What the analyst needs is not clairvoyance but assembly.

There is a hard boundary here, and it is worth stating plainly rather than piously. Trading on material non-public information is a crime; building a conclusion from dispersed public sources is the job. We work from filings, public statements and cross-chain inference, and we stay away from the conversations most likely to blur the two, which has the convenient effect of removing the temptation before it arrives.

What results is strange enough to deserve a name. Piece together a hyperscaler's capital expenditure from the published expansion plans of its optical transceiver suppliers, and you hold something functionally new to the market, though not one component of it was hidden from anybody. Nobody withheld it. Nobody ignored it. It simply never existed in a single place until someone put it there.

A question that often comes to our minds: we can easily see how our work on our first fund themes set the base for the second, but does the work we do for the new fund have value for investors in the first? And, the conclusion so far appears clear: the work on the second fund themes and stocks tremendously add to our understanding of evidence supporting the first fund theses. 

Pioneers of Our Own Ignorance and Arrogance

Now the harder direction. Everything above concerns what the world knows. The more awkward interrogation runs the other way.

We can name, without much effort, a dozen ideas over three years where we felt we were first: the hardware-to-software inversion, instant copyability, silicon scarcity, token inequality, our early insistence on memory. While we are proud to see what we deem as “our” themes discussed by others, we at times feel hard done by if we observe someone else claiming the idea as theirs and benefiting more. 

The situation that stung us the most was when we saw a 164-page paper published in 2024 that aroused everyone’s awareness. Until recently, we privately believed we had been a year ahead of most themes discussed in that paper, and the paper mostly spoke our language in the articles timestamped and available on our website. Of course, our envy was most for the billions the entity could use on the same ideas while we were struggling to justify our inexperience and extremeness. Given what has since become of that venture, we feel more comfortable airing our grievance. But this schedenfraude often makes us wonder who are truly idea pioneers and who truly have claims to any idea discovery. 

The correction, in our case, arrives whenever we search earnestly about the ideas we thought were ours for evidence of others holding similar thoughts. Invariably, we meet a world where every claimed original thought seems to have many priors. In our case, we read our minds through articles expressing views not just far earlier than our publications but often far better expressed. This is the ordinary condition of ideas rather than a humiliation even for genuine pioneers. Newton and Leibniz. Darwin and Wallace. Bell and Gray, filing at the same patent office on the same day.

An empty space on our own map feels like unexplored territory for the species. It is almost always just an empty space on our map.

How Deep Is Deep Enough?

There is no floor to what one can learn about any of this. For most investors, knowledge of advanced packaging runs precisely as deep as the company's slide: a handful of diagrams, three acronyms, faithfully recycled by analysts who received them the same way. Open the journals, the teardowns, the process papers, and a subject a thousand times larger appears, in which the diagrams are not so much wrong as decorative.

Every technology has an ocean beneath the slide, and a financial investor cannot drink oceans. The new fund has turned this from an occasional problem into a daily one, since it introduces an unfamiliar subsector roughly every week, each with its own ocean and its own reassuring diagram.

So the question is not how much can be known but how much needs to be. Part of the honest answer is competitive pride: we want to know slightly more than the manager in the office next building, and slightly more is a wonderfully renewable ambition. The useful answer concerns relevance, and the point at which the next layer of detail stops changing the conclusion.

Knowing more than the market is an edge. Knowing more than the decision requires is a hobby. Both feel identical while you are doing them, but neither might be beneficiary for investments in the fund.

The Ledger of Amnesia

In a world of constant stimulation and constantly arriving information, even those born with the best memories are often unaware of what they actually know. And this writer, as anyone who has watched me fail to place a familiar face will confirm, is not blessed in things called memory. 

What we consider new information is often forgotten information. The world itself forgets what it knows and calls the rediscovery new. The analyst's version is the “new" thesis that was in your own notes three years ago. You read a research note in March. You file it. By August, the same thesis reappears in a different format — a new headline, a fresh chart, a CEO quote you have never heard — and you feel the old shiver of discovery. It is not discovery. It is delayed receipt. The information was already in your possession; you had simply lost the index. The mind treats forgotten knowledge as new knowledge, and awards itself the same dopamine prize.

The worst variant is the forgotten position. You buy a stock for reasons that seemed urgent at the time. Six months later, the price has moved, and you cannot remember why you own it. You reconstruct a thesis from the current price action, as if the market is telling you what you once knew. It is not. The market is telling you what everyone else now knows. And you are listening as if it were a revelation.

Known to the Agent, Unknown to Us

A team has never known the sum of what its members know. That is not a failure of character. It is what happens when four people read different things and meet on Tuesdays. A good share of the institutional disasters of the past twenty years were foreseen by somebody in the building whose memo reached a folder rather than a decision.

We have now added a stranger layer by outsourcing part of remembering. Somewhere in our systems sit thousands of pages that were ingested, classified, summarised and stored, and read by no human on the team.

Is that known to us? A serious argument says yes: it was processed, ranked, retained, and any of us could retrieve it in seconds by thinking to ask. An equally serious argument says that knowing has never meant storage. Knowing means the thing surfaces uninvited at the moment it becomes material, and materiality arrives long after the filing does. 

Which leaves a failure mode unavailable to earlier generations of investors. It is now entirely possible for a firm to possess a fact, to have summarised it competently, to have filed it accurately, and to be blindsided by it.

In Conclusion: Whom Does One Believe?

This could meander for another ten thousand words. The masters of literature built careers on the follies of memory and understanding, and they never had to set one person's follies against the average of everybody else's.

So, one last question in place of an ending. A commentator declares that the operators are torching capital, wielding ROI like a curse word, producing a chart of the Nasdaq between 2000 and 2002 and two paragraphs on the South Sea Bubble. Is his synthesis the one to trust?

Consider the alternative. The operator has read the same history. He is asked about it in board meetings, on earnings calls, by journalists, at town halls, and quite probably across his own dinner table. He carries the doubt continuously and is made to answer it publicly. And he is still committing tens of billions in the direction that is harder to defend, on the strength of evidence he examines daily, and the man holding the history book has never once opened.

Either of them may be wrong. That was never the point. Believing anyone requires more than collecting diverse opinions. It requires asking what each of them paid for what they claim to know, in time, in attention, and in the risk of being wrong in public. However, it is not just about who has what skin in the game or motivations, but also about who has access to information the other does not have.

Enough philosophical mumbo-jumbo. There is a great deal of work pending on GenInnov Epicentre.