A100 Is the Dress Rehearsal

AI stops being a laboratory project and becomes institutional capex.

In May 2020, the world is distracted by a pandemic, strained supply chains, empty offices, markets distorted by monetary policy, and companies trying to digitize in months what they had avoided for years. Amid that noise, accelerated computing takes another step. The A100 should not be read only as a new GPU. It should be read as the dress rehearsal for an era in which AI, analytics, training, and inference begin competing for the same infrastructure budget.

The market likes to divide technology into clean phases. First research, then product, then adoption, then maturity. Reality is messier. Research, product, competitive fear, excess capital, and infrastructure advance together. Sometimes the final product is not yet clear, but the race for capacity has already begun. Large companies do not buy infrastructure only because they know exactly what they will do with it. They also buy because they do not know what the competitor will do.

A100 represents a category of hardware designed to unify workloads that previously seemed separate: training, inference, data analytics, HPC. The technical detail matters, but the economic reading matters more. The more a platform can serve multiple workloads, the greater the chance it enters the budget as a strategic asset. A chip that is too specialized can remain trapped in a niche. A sufficiently versatile accelerator can become a data center layer.

Nvidia and AMD are the obvious names. Nvidia because it leads the narrative and ecosystem of accelerated computing. AMD because every large customer hates absolute dependence on one supplier, and because data centers will seek alternatives whenever performance, price, availability, or negotiation allow. But the less obvious names show where the expansion becomes structural: Broadcom, Arista, Marvell, Credo.

Broadcom captures pieces of connectivity, networking, infrastructure silicon, and components that become more valuable as data centers become machines of internal traffic. Arista captures high-performance networking and automation in cloud environments. Marvell appears in silicon for infrastructure, connectivity, storage, and data center. Credo, still smaller and less known, represents the thesis of high-speed interconnect, active cables, and solutions for moving data efficiently.

The investor needs to understand one simple thing: AI at scale is a war against waiting. Waiting to train. Waiting to move data. Waiting to fetch memory. Waiting to synchronize machines. Waiting to deliver a response to the user. Waiting is cost. Companies that reduce waiting capture value.

Perhaps in 2022, with a new generation of accelerators and a more mature narrative around generative AI, the market begins to see that the race is not only for better chips, but for entire systems. H100, or any later generation, would be the psychological moment when the industry stops talking about experiments and starts planning capex. The difference is enormous. Experiments fit inside innovation budgets. Institutional capex enters strategic planning.

The way to profit is to observe the transition between curiosity and obligation. While AI is curiosity, companies buy little, test, hire consultants, run pilots, and publish press releases. When it becomes competitive obligation, they buy capacity before knowing the exact ROI. That phase is dangerous for application investors, but powerful for infrastructure suppliers. Whoever sells the pickaxe does not need every miner to find gold. He needs many to dig.

Broadcom and Marvell can capture value in the chip and connectivity layer. Arista can capture value in network architecture. Credo can capture value where interconnection between components needs speed and efficiency. Nvidia captures the emotional center of the race. AMD tries to capture the need for an alternative. The question is not who has the most beautiful story. It is where the constraint tightens first.

The counter-thesis is severe. A100 can generate enthusiasm before sustainable demand. H100 can become a symbol that is too expensive. Companies can overbuy and then pause. Applications can fail to monetize. The networking chain can face competition and commoditization. Credo can be too small and too volatile. Marvell can promise data center exposure and deliver a mixture of cycles. Arista can be excellent and still become expensive. Broadcom can be more conglomerate than pure thesis. AMD can remain behind. Nvidia can be a great company and a bad stock if bought without price.

But the investor does not need to deny these risks. He needs to understand the mechanism. When a computing platform becomes the basis of competition, customers buy capacity in waves. The first wave is technical. The second is strategic. The third is defensive. The defensive wave is usually the most irrational and the most profitable for suppliers.

In May 2020, no one needs to know exactly which final application will win. That is the beauty of the thesis. Perhaps language models. Perhaps recommendation. Perhaps computer vision. Perhaps drug discovery. Perhaps security. Perhaps customer-service automation. Perhaps assisted programming. Perhaps industrial analytics. The application may still be uncertain. The need to accelerate experiments is less uncertain.

The investor who demands total clarity on the final product usually pays too much when clarity arrives. The price of clarity is lower return. The premium belongs to whoever accepts technical uncertainty with the correct economic structure.

A100 is the dress rehearsal because it shows the direction: data centers will not be only CPU warehouses. They will be heterogeneous factories, with accelerators, fast networks, specialized memory, and orchestration software. Computing will stop being a straight line and become an architecture of multiple engines. Each engine will have a supplier, margin, cycle, and bottleneck.

The market is rehearsing calling this AI.

I would call it the industrialization of calculation.

Leo Bentier