News vs. Noise: Scarce Hardware and Data Control Are Driving AI’s Next Phase

Software Efficiency Is Fueling Demand, Not Shrinking It

The conventional worry about AI has been that as models get more efficient, the industry will need less hardware to run them. That assumption broke down this week. At Google’s 2026 I/O keynote, the company confirmed that token processing volume, essentially the raw amount of AI computation happening across its systems, grew sevenfold over the past year [1], far outpacing the roughly 30% efficiency gains that analysts had already built into their demand forecasts [2].

What’s actually happening is a pattern economists call the Jevons Paradox: when a resource becomes more efficient to use, people don’t use less of it; they find more ways to use it. Every unit of memory freed up by smarter software is immediately reinvested in longer context windows and more AI agents running in parallel.

That keeps pressure on High Bandwidth Memory (HBM), a type of high-speed memory chip that stacks layers of DRAM to enable AI processors to move data faster while using less power. The numbers make the shift concrete: HBM content jumped from 80GB in Nvidia’s H100 chip to 372GB in the newer Blackwell GB200, a 4.6x increase [3], [2]. Structurally, that means every new generation of AI server needs far more of a component that’s already in short supply [3], [2]. For investors, this points to the physical layer of the AI buildout being a durable trend rather than a passing spike, though it’s worth noting that HBM investments carry real risks, including fast-moving technology cycles and heavy upfront capital costs [2].

The AI Trade Is Moving From Rented Intelligence to Owned Infrastructure

Large technology companies have spent the past few years renting AI capability through APIs, the standardized connections that let one piece of software request services from another over the internet. That’s starting to change. Companies are increasingly building or buying their own AI infrastructure instead of leasing it [4].
The driver is data governance, the set of rules and systems a company uses to keep its data secure, private, and accurate [4]. Alibaba’s July 3 decision to ban employees from using outside tools like Anthropic’s Claude Code, in favor of its own internal platform, is a clear signal of this shift toward what’s known as sovereign AI, where a company keeps its data and models entirely inside its own walls [5], [4]. A June partnership between Palantir and Nvidia reinforced the same trend, giving enterprises a way to run AI models locally rather than through a third-party cloud [6], [4].
For investors, the opportunity is migrating toward what you might call the control layer: identity management systems that verify who or what is allowed to access a company’s AI models and data, and the private infrastructure that supports them [4]. As with any fast-evolving corner of enterprise technology, this space carries risks, including rapid shifts in underlying technology and stiff competition from major cloud providers [4].

Brain-Inspired Chips Are Leaving the Lab

Power consumption is the main obstacle to AI in small devices like drones and wearables [7]. That’s the problem neuromorphic computing is built to solve: a chip architecture modeled loosely on the human brain, designed to be event-driven and energy-efficient rather than running constantly at full power [7].
Intel’s Hala Point, a research prototype with 1.15 billion artificial neurons, has already proven that the architecture can scale [8], [7]. This week, BrainChip’s Akida processor marked something more significant: the jump from lab prototype to a functional, revenue-generating product [9], [7]. That transition matters because it’s rare, making BrainChip one of the few pure-play options in the space, meaning the company’s entire business is built around this one technology rather than treating it as a side project buried inside a larger, diversified business [7], [9]. For an investor curious about direct exposure to this theme, that distinction is worth understanding, alongside the risks that come with any early-stage chip architecture: high research costs and thinner trading liquidity than you’d find in a large, established company [7].

Samsung’s Stock Drop Is About Leverage, Not Memory Demand

Samsung shares fell 9.8% in Seoul on July 10 [10], [11], even after the company forecast a 19-fold jump in operating profit, the amount left over after a company pays its core business costs [10]. On the surface, that looks like a contradiction worth worrying about.
It isn’t. The selloff reflects forced deleveraging, where institutional investors who borrowed heavily to build a position get forced to sell quickly to reduce that borrowed exposure, combined with a broader unwind of a crowded trade, meaning too many investors held the same bet at once [12], [11]. When the selling started, there weren’t enough buyers on the other side to absorb it without pushing the price down sharply [12], [11]. Neither of those dynamics says anything about whether the world actually needs less memory [2]. A one-day price swing in an international market rarely overrides the reality of a trillion-dollar, multi-year wave of AI infrastructure spending already underway [2].

A Celebrity Short Bet Isn’t a Portfolio Strategy

Reports that “Big Short” investor Michael Burry has taken a bearish position against Micron, meaning he’s borrowed shares to sell now with plans to buy them back later at a lower price, made headlines this week [13]. It’s easy to read that as a signal.

It’s a weaker one than it looks. Following any single investor’s trade, however famous, skips the transparency that a full portfolio strategy requires, and market timing bets rarely translate cleanly into a plan you can actually follow [14]. High-profile short positions are also frequently part of hedging structures that are never disclosed publicly, meaning the trade might exist to offset risk elsewhere in that investor’s book rather than representing a standalone view that Micron is doomed [14].