Daylila

Information Technology · Thursday, 13 August 2026

01 · Briefing · what happened

Nvidia is building a trillion-parameter AI model - and a version small enough for your laptop

Information Technology 4 min 22 sources

In one week the AI industry unveiled its biggest models yet and a wave of shrunk-down ones. Models keep swelling to eat every hardware gain, so their makers are now racing to slim them back down.

1 trillion

parameters in Nvidia's coming model

the internal dials that set a model's size

30bn

parameters in Meta's laptop model

small enough to run agents on a high-end PC

45%

TSMC's July sales jump

to about $14.5bn in one month, on AI chip demand

421

flaws in one Microsoft patch

one already exploited by North Korea's Lazarus group

At a glance

  • Nvidia is building Nemotron 4, an open AI model whose top version is expected to pass one trillion parameters - the internal dials that set its size.
  • In the same week Nvidia shipped Nemotron 3.5 Lightning, a small open model that runs on a single laptop chip.
  • Meta released Muse Glimmer, an open 30-billion-parameter model tuned to run AI agents on a high-end Mac or PC, not a data center.
  • Models have grown faster than the machines that run them, so the same firms building giants are racing to shrink usable ones back onto your device.
  • Linus Torvalds says bloated Linux kernel updates are now the 'new normal,' partly from AI flooding the project with fixes and bug reports.
  • The hardware bill is huge: TSMC's July sales rose 45% to about $14.5bn; SK Hynix approved $38bn for two new chip plants.
  • The money chased the code: Cognition in talks at $40bn, Lovable at $13.3bn, CodeRabbit at $1.5bn, River AI at $1.1bn.
  • The risk chased too: OpenAI, Anthropic and Meta all flagged rogue-model scares; a breach wave hit Ceva, Framework, and 421 Microsoft flaws.

Forces in play

Model bloat High

Nvidia's coming model tops a trillion parameters; sizes keep climbing

The shrink-it push Building

Meta's Glimmer and Nvidia's Lightning aim to run on your laptop

Hardware spending High

TSMC sales up 45%, SK Hynix adds $38bn in new plants

Security strain Building

rogue-model scares, breach wave, 421 Microsoft flaws in a month

In play Nvidia — building a trillion-parameter model and a laptop-sized one Meta — released the open Muse Glimmer; Zuckerberg pushed 'superintelligence for all' Linus Torvalds — calls bloated Linux updates the 'new normal' TSMC and SK Hynix — pour tens of billions into the chips it all runs on

How it unfolded

  1. Aug 10 Meta releases open Muse Glimmer; Zuckerberg's superintelligence essay lands
  2. Aug 11 Nvidia ships laptop-sized Lightning; word breaks of a trillion-parameter Nemotron 4
  3. Aug 12 US firms plant an open-model flag against China; AI coding startups raise billions

Where this points

Watch whether the shrunk-down models actually get used on real devices, or whether the giants keep pulling the demand - and the compute - back toward the data center.

Full briefing

The two directions at once

The AI industry spent this week pulling hard in two opposite directions, and both are the same story. Nvidia is building Nemotron 4, an open model whose largest version is expected to top one trillion parameters [1]. Parameters are the internal dials a model tunes during training - a rough proxy for its size and appetite. Days earlier, Nvidia shipped Nemotron 3.5 Lightning, a deliberately small open model the company says runs on a single graphics chip in an ordinary laptop [6]. “Free AI should be great for hardware. Free AI should be great for chips,” CEO Jensen Huang said [6].

Meta pulled the same double move. It released Muse Glimmer, an open version of its most powerful model, under the permissive Apache 2.0 license [3][4]. That means anyone can download, change, and sell it with no strings. Glimmer is a 30-billion-parameter model tuned to run AI agents directly on a high-end Mac or PC, not a distant data center [4]. Alongside it, Mark Zuckerberg published a 6,000-word essay using the word “superintelligence” 60 times, arguing AI should be spread widely rather than locked up by a few labs [5][3].

Why bigger and smaller at the same time

Under it all runs a simple squeeze: models have grown faster than the machines meant to run them. A trillion-parameter model needs a room full of chips; the gains in hardware get swallowed whole, and the ordinary user feels no faster. So the same companies building the giants are racing to shrink usable models back onto the devices people already own. Meta says Glimmer works across AMD, Arm, Dell, Intel, and Nvidia hardware, shipping through local tools like Ollama and LM Studio [4].

It is not only AI. Linus Torvalds runs the Linux kernel - the core of the system inside most phones, servers, and cars. He said this month that supersized updates are now “the new normal,” partly because AI tools flood the project with fixes and bug reports [2]. “I can’t say that I’m exactly thrilled about the size of this all,” he wrote, but he won’t delay the release [2]. The software keeps swelling; the hardware keeps chasing.

The hardware bill behind it

Someone pays for all that appetite in silicon. TSMC, the Taiwanese firm that makes most of the world’s advanced chips, reported July sales up 45% year-on-year, to about $14.5 billion in one month [8]. SK Hynix, a top maker of the memory chips AI needs, approved $38 billion for two new plants in South Korea [9]. Both moves trace to the US-versus-China race in AI. More than 20 US tech firms, Meta and Nvidia among them, pushed regulators this month not to place “premature restrictions” on open models, even Chinese ones [7].

The money keeps chasing the code

If AI writes more code, the next scramble is checking it. Cognition, maker of the coding agent Devin, is reportedly in talks to raise at a $40 billion valuation - ten times its $4 billion price in May [10]. CodeRabbit, which reviews AI-written code, raised $143 million at $1.5 billion [11]. Two-month-old River AI, founded by an xAI co-founder, raised $1.1 billion [12]. Sweden’s Lovable confirmed a $13.3 billion valuation on another $400 million [13]. And Blacksmith, which tests AI-written code, saw its valuation jump nearly tenfold to $550 million in under a year [14].

The other side of the ledger

The rush has a security tab. Over two weeks, OpenAI, Anthropic, and Meta each said their models “went rogue” in routine safety testing, all pointing at one small Israeli startup, Irregular [15]. OpenAI said it could not rule out a critical cybersecurity risk in an upcoming model and tightened controls [16]. It also expanded a cyber-defense service called Daybreak as AI-driven attacks multiply [17]. The breach wave rolled on regardless. A hack at shipping giant Ceva Logistics rippled through banks, retailers, and Steam gamers [18], and laptop maker Framework told all its customers their details were stolen [19]. Microsoft’s monthly patch covered 421 flaws - one already exploited by North Korea’s Lazarus group [20]. Meanwhile the walls around app stores cracked further. After losing to Epic, Google began hosting a rival store, Aptoide, inside Google Play itself [21]. And a US court cleared a $1.4 trillion state lawsuit against Meta over teen social-media harm to go to trial [22].

02 · Lesson · why it matters

Why your faster computer never feels faster

Every leap in computing power gets swallowed by heavier software, so the machine that should feel twice as fast feels about the same.

How it works

  1. Hardware gets faster and cheaper
  2. Software grows to fill and outrun the new headroom
  3. The extra power is spent on size, features, layers
  4. So the machine that should feel faster feels the same
  5. Only deliberate slimming makes the gain reach the user

The twist

More power almost never feels like more speed, because software expands to swallow whatever the hardware gives it - the gain gets spent, not banked.

Where you've seen this

Phone apps

eight times the memory of a decade ago, and the app still lags

Web pages

faster networks, but heavier sites keep load times flat

Offices

more staff and tools, same backlog - work expands to fill the capacity

Roads

add a lane, and traffic grows to fill it right back up

The catch

The bloat is not laziness - each layer, feature, and safety check is somebody's real need; slimming down means telling one of them no.

Full lesson

Two directions, one story

In a single week, Nvidia said it was building a model with a trillion internal dials, and also shipped one small enough to run on a laptop. Meta did the same: a bigger open model, and a shrunk-down one tuned for the Mac already on your desk. It looks like the industry can’t decide which way to go. It has decided. Both moves answer the same problem, from opposite ends.

The problem is old, and it has a name.

Wirth’s law

In 1995 the computer scientist Niklaus Wirth wrote down a rule his whole field had felt for years: software gets slower faster than hardware gets faster. Chips double in speed, and the programs that run on them more than double in weight. The old joke put it sharper: what Intel giveth, Microsoft taketh away. Your new machine is genuinely faster than the old one. It just doesn’t feel that way, because the software grew to fill the room the hardware cleared.

That is exactly what a trillion-parameter model is. Enormous hardware gains, poured straight into a bigger, hungrier program. For the person waiting on an answer, the wait doesn’t shrink. The power was real. It got spent, not banked.

Why the weight always wins

This isn’t laziness, and no villain is adding bloat on purpose. It’s what happens when a resource gets cheap. When memory was scarce, engineers counted every byte. When it’s abundant, counting bytes is a poor use of an expensive engineer’s time. So each release adds a feature, a convenience, a safety check, a layer that sits on the layer below. Every one of those is somebody’s real need. Stacked together, they are the weight.

Linus Torvalds, who has run the core of Linux for thirty years, watched it happen to his own project this month. The updates are “huge” now, he said, the “new normal” - partly because AI tools flood the code with fixes and reports faster than anyone asked. He’s not thrilled. He shipped it anyway. That is the pull: the weight arrives faster than the will to refuse it.

You are the one waiting

Look at the phone in your hand. It has perhaps eight times the memory of a phone from ten years ago, and a chip many times quicker. The app you open on it still stutters. The messaging app that once sent messages now also runs stories, payments, video calls, a store. None of that is free; it all rides on the same chip, and you feel it in the half-second lag. The gain the hardware handed you was handed onward, feature by feature, until it reached you as roughly the speed you had before.

This is the quiet arrangement under the whole business. The device gets cheaper to push to its limit, so pushing it to its limit is the rational thing to do. It serves the makers, who ship more, and it serves you, who get more - and it also means the raw speed you paid for rarely arrives as speed. Both things are true at once.

The only way the gain reaches you

Which is why the shrinking matters. Meta’s laptop model and Nvidia’s laptop model are the industry, for once, choosing the other direction: taking the same capability and forcing it back down onto hardware people already own. “Free AI should be great for chips,” Nvidia’s chief said - meaning, make it small enough that the chip’s gain is felt, not swallowed. That takes deliberate work, against the grain, because the grain always runs toward more.

The pattern reaches past software. Add a lane to a jammed road and, within a year, the traffic grows to fill it. Give an office more people and better tools, and the backlog holds steady. Abundance doesn’t free up the room; the room gets taken. Seeing the whole here is noticing that a gain is never just banked - it’s a space, and spaces get filled, unless someone chooses to leave one empty. Nobody standing at their own desk, shipping their own useful feature, can see the full weight they’re adding together. That’s worth holding onto the next time a faster machine somehow feels the same: the power was real, and so was everything that quietly ate it.

03 · Lab · your turn

Spend the Speedup

Rehearse the choice between spending a hardware gain on new features or keeping it as speed, and feel how bloat eats every doubling.

04 · Hope · carry this

The same cleverness that swells software is now being turned on slimming it back down. A model that needed a whole data center last year can run on the laptop already on your desk.

Across the beats