NVIDIA: The Video Card Is a Calculation Machine Wearing a Toy Costume

The market likes childish names for things it does not yet understand.

It calls Nvidia a videogame company. It calls the GPU a graphics card. It calls the end user a gamer. It calls the product an accessory. Then it closes the subject, as if classification were analysis. It is not. Classification is often the tomb where Wall Street buries its own curiosity.

Today Nvidia is executing a 3-for-2 split. There is no alchemy in this. Three smaller pieces are not worth more than two larger pieces. The split does not produce cash, increase margins, create customers, improve architecture, pay debt, or transform a bad business into a good one. A split is arithmetic cosmetics. It is the company telling the market: "the price rose too much to look small, so we will make it look smaller."

But sometimes cosmetics reveal traction.

Dead companies do not split to improve liquidity. Irrelevant companies do not need to look accessible. The split, by itself, is not the thesis. That would be amateur thinking. The thesis is in the reason the split became possible. And the reason appears to be hidden in the most obvious place: the GPU.

The GPU was sold to the world as a machine for pushing pixels. But pixels are not toys. Pixels are matrices. Matrices are mathematics. Mathematics is calculation. And calculation, when multiplied in parallel, stops being visual decoration and becomes infrastructure.

The gamer believes he bought a card to kill monsters on a screen. Perhaps, without knowing it, he bought a piece of the future calculation machine of the world.

This is how great asymmetries appear. They enter through the back door, dressed as something vulgar. Radio was entertainment. The telephone was a curiosity. The personal computer was a nerd toy. The internet was a catalog. The smartphone was a phone with a screen. The GPU, they now say, is for videogames. This kind of sentence tends to age badly.

The CPU thinks like a bureaucrat: one thing after another, order, sequence, queue, stamp. The GPU thinks like a crowd: thousands of small similar operations happening at the same time. To draw a scene, it must calculate light, texture, position, shadow, movement, depth. All of this is parallel calculation disguised as image.

The market is looking at the disguise.

I am not interested in the game. I am interested in the engine of the game. I do not care about the teenager buying the card to get more frames per second. I care that this teenager, with his household budget and visual addiction, is subsidizing a chip production chain that may later serve scientists, engineers, laboratories, banks, governments, energy companies, pharmaceutical companies, and, at the limit, machines that learn.

The economic genius of this is obscene. The consumer funds scale. Scale lowers cost. Low cost attracts new uses. New uses demand software. Software traps developers. Developers create dependency. Dependency creates ecosystem. Ecosystem creates margin. Margin creates power. And the market, late as always, will call it "inevitable" when it is already too expensive.

There is a word beginning to matter: CUDA.

It is not beautiful. It sounds like an engineer's tool. Precisely for that reason it deserves attention. The market likes slogans. I prefer ugly infrastructure. CUDA may be the bridge between hardware that already exists and applications that do not yet fit inside analysts' vocabulary. A powerful GPU without a programming layer is a power plant without wires. The electricity is there, but few can use it. CUDA is an attempt to turn raw power into a programmable tool.

This changes the nature of Nvidia.

Without CUDA, Nvidia sells cards. With CUDA, Nvidia begins to sell a way of thinking about computation. The difference between selling a tool and selling a language is the difference between selling bricks and defining the architecture of the city.

The lazy investor will ask: "How many games will be sold?" The correct question may be: "How many of the world's problems are parallelizable?"

Physical simulations are parallelizable. Molecular modeling is parallelizable. Medical imaging is parallelizable. Financial risk is parallelizable. Seismic research is parallelizable. Scientific computing is parallelizable. A significant part of artificial intelligence, if it ever stops being an academic promise and becomes an industry, will be parallelizable.

The obvious objection is that this is still small. Good. Small is the stage before large. The obvious is merely the late with self-esteem.

Another objection: Intel dominates computing. Yes. And that is exactly why the opportunity exists. Consensus sees only the throne. Asymmetry sees the siege. The CPU will not disappear. That is a childish reading. The world does not need to exchange one religion for another. It needs to move certain workloads to the kind of machine that executes them better. When the problem is sequential, use CPU. When the problem is parallel, use GPU. The question is not total replacement. The question is incremental capture of a new bottleneck.

The market pays for categories. I prefer buying transitions between categories.

Nvidia is still mentally priced as a graphics company. But what is being built, perhaps without the market accepting it yet, is a parallel-computing platform. This kind of change rarely appears first in profit. It appears first in small anomalies: researchers using gaming cards for science; developers learning a new layer; companies testing GPU clusters; engineers discovering that a part made to render fantasy can accelerate real work.

The modern world has a simple problem: it wants to calculate more than the old architecture allows at an acceptable price.

There are two ways to solve this. The first is to build expensive centralized monsters, revered by universities and governments. The second is to use components mass-produced for an apparently vulgar industry and redirect them toward serious problems. The second way is uglier. That is why it interests me.

Beauty is usually expensive. Ugliness is often mispriced.

Nvidia is in a strange position. It depends on games, but may not be defined by games. It sells hardware, but may be digging a software moat. It competes in semiconductors, but may be entering the platform economy. It looks cyclical, but may be planting something secular. And this kind of ambiguity is exactly where the market makes large mistakes.

I am not saying it will be clean. Nothing worth money is clean. There will be inventory cycles. There will be competition. There will be multiple compression. There will be analysts saying margins have peaked. There will be bad quarters. There will be the temptation to sell while the thesis is still maturing. The investor who needs quarterly comfort should not buy asymmetries. He should buy consensus and complain about the return.

The question is not whether Nvidia is cheap on every spreadsheet metric. The question is whether the market is using the wrong spreadsheet.

If you value a bridge as a pile of concrete, it always looks expensive. If you value a calculation platform as a maker of boards for teenagers, it may look expensive before it looks ridiculously cheap.

How would I position?

Not with heroism. Heroism in markets is only leverage wearing makeup. I would buy the classification error, not the euphoria. The correct position would be a long equity position, sized to survive a large decline without forcing liquidation. The point here is not to guess the next quarter. The point is to own a long option on the possibility that Nvidia is reclassified from a games company to computing infrastructure.

I could complement with long calls, if the pricing structure were sufficiently cheap, but only as a satellite, never as the core. Options are useful instruments when they buy convexity. They become poison when they buy vanity. The core would be equity. Common stock does not expire. The thesis may need years to be understood. The market may take longer than the common speculator's patience. That is why the instrument must breathe.

I would not use margin. Margin turns a right thesis at the wrong time into certain ruin. Nvidia can fall 40% without the thesis being dead. It can fall more if the semiconductor cycle turns, if the market corrects, if growth appears to slow, if Intel or AMD promise to kill the party. The investor who cannot withstand volatility does not own a thesis. He owns only a quote.

Nor would I buy all at once. Arrogance likes a single entry. Prudence prefers building a position. I would buy an initial piece to force myself to follow the company with real capital, not cheap opinion. I would add if signs of non-graphics use kept appearing: CUDA adoption, scientific applications, clusters, partnerships, developers, evidence that the GPU is leaving the periphery and becoming an engine.

The main risk is not competition. Competition exists in every good thesis. The main risk is that the market is right and the GPU remains merely a piece of the graphics cycle, without relevant economic capture outside games. In that case, the split will be only a split, CUDA will be a curiosity, Tesla will be marketing, and Nvidia will be another semiconductor company with a good story and a dangerous multiple.

But there is a detail: some dangerous stories are dangerous because they are false. Others are dangerous because they are too early.

I suspect this one belongs to the second family.

The average investor wants proof before paying. But when proof arrives, the price has already changed religion. The big money is rarely in accepting the obvious. It is in tolerating the ridiculous before it is promoted to consensus.

Today, in 2007, calling Nvidia essential computing infrastructure would sound exaggerated. Good. Exaggeration is the tax paid by anyone trying to see a change before it becomes an investment-bank slide.

Perhaps Nvidia will continue to be seen as a video-card company. Perhaps the market will take years to understand that a video card is a calculation machine wearing a toy costume.

But if the costume falls, the old multiple falls with it.

And when Wall Street discovers that it was not looking at a toy, but at a new form of computation, it will do what it always does: pretend it knew from the beginning.

Leo Bentier