Gerry Burde

[Gerry Burde @GerryBurde]

@TechFundies Same reason all the labs are building their own inference chips. keep the margin for themselves. Also look at the market caps.. Meta is $2T, with $70B in profit... Meanwhile OpenAI and Anthropic are both $2T with no profit. They could double their cap if they make SOTA models

[Gerry Burde @GerryBurde]

@GregHBurnham I don't know, I have for all intents unlimited Opus 5.5 / Astra 6 instances in 2026 and it hasn't made me rich yet. One instance would be a bit limiting, but at least a few million by 2020.

[Gerry Burde @GerryBurde]

People are bringing 10 year old GPUs back online for inference... Newer more efficient GPUs aren't replacing older ones.

The "token factory bankruptcy" narrative applies a 19th-century railroad bust or merchant-power model to an asset class operating under entirely different unit economics, capital structures, and physical constraints.

To see why a systemic collapse in token production isn’t materializing, you have to break down three structural realities:

1. Marginal Costs Are Not Stranded by Power or Facilities
In traditional commodity collapses (like natural gas or merchant power plants), operations halt because marginal cash operating expenses (fuel, facility rent, turbine maintenance) eclipse the market clearing price.

Inference doesn't work that way. Datacenter rack space, cooling, and wholesale electricity account for a single-digit percentage of total token delivery cost. When the marginal cost to keep a chip powered is practically negligible compared to its capital cost:

Operators don't shut chips off when prices drop—they bid them into lower-tier inference tasks.

Even decade-old hardware (such as legacy Pascals or early Tensor Core architectures, V100's etc) is actively racked and profitable today for low-throughput, batch, or non-latency-critical serving.

There is no operational cash drain forcing operators into Chapter 11 shutdowns simply because token spot prices dip.

2. Balance Sheets Are Captive, Not Levered Merchant Foundries
A traditional capex bubble pops when independent producers finance hardware with high-yield debt, banking on fixed forward contracts that get broken.

The vast majority of hyperscale token production is funded directly out of the balance sheets of cash-flow machines—enterprise software, search monopolies, and advertising platforms running 30–40% operating margins:

These entities do not face bank-driven liquidation cycles.

If hardware becomes economically obsolete before its scheduled accounting depreciation, it gets written down on the balance sheet as an impairment, not liquidated in a bankruptcy auction.

Captive internal demand (search enhancement, coding assistants, internal tooling, ad targeting models) absorbs capacity even if external API prices crater.

3. Autoregressive Inference Is a Memory Wall, Not a Compute Glut
The core premise of the commodity-bro thesis is that flooded silicon capacity drives token value to zero. But inference is not compute-bound; it is strictly memory-bandwidth bound.

The Generation Chokepoint: During autoregressive generation (token-by-token decoding), the model must stream entire weight matrices through memory just to produce a single token (arithmetic intensity near 1 FLOP/byte).

The HBM Moat: A glut of raw tensor FLOPS does not crash token supply if memory bandwidth remains scarce. High Bandwidth Memory (HBM) packaging and interconnect fabric are the actual bottleneck.

The Floor on VRAM: Because inference throughput is dictated by gigabytes-per-second of bandwidth rather than raw compute cycles, older high-bandwidth silicon continues to command a bid. Every unit of memory bandwidth that can hold model weights and the KV cache generates revenue.

A systemic bust would require an order-of-magnitude architectural disruption in memory bandwidth efficiency—such as radical sparsification, 1-bit quantization without perplexity penalties, or non-autoregressive architectures that break the memory-streaming constraint entirely. Until that physical bottleneck disappears, memory scarcity establishes a structural pricing floor that prevents an unconstrained supply collapse.

[Gerry Burde @GerryBurde]

@_FORAB There will be nearly unlimited demand for DDR5 if prices for it were back at 2024 levels. We don't need HBM for inference, it's just a nice to have. The incumbents should be scaling all of their manufacturing to keep market share, but CAPEX barely seems to have accelerated all

[Gerry Burde @GerryBurde]

It's not because it's hot, it's because 62,000MW of capacity is offline for maintenance(which is scheduled for this cooler period of time..). This notice isn't an actual emergency, it's just the method used to allow heavy industrial users to switch to their on site power without constraints

[Gerry Burde @GerryBurde]

If Maia is in fact 1.7x more efficient than GPU, and they produce them for cheaper than the NVIDIA tax, I'm sure it'll reduce demand for private inference by the fortune 500 and other companies/CoPilot users that trust Microsoft or have other reasons to pay a premium for hosted inference. Eventually the default should become private inference, but the TPU's or other solutions need to actually be available and not in a walled garden of hyperscalers

[Gerry Burde @GerryBurde]

@LeachartsMarte @HannoKlausmeier What makes it expensive? Batteries? Have you considered they will be 90% cheaper in 5 years? Electricity? Have you considered you can make your own for 5c/kwh in the back 40? Is it the reduced maintenance?

[Gerry Burde @GerryBurde]

@BridgeObsidian @satyanadella Hyperscalers wouldn't, but someone should be building these for the several trillion dollar enterprise and consumer market.

[Gerry Burde @GerryBurde]

It's most likely the opposite. More people will go to restaurants because the quality will(can) be higher due to shorter inventory cycles, it can be more varied due to larger inventory of spices and ingredients, higher quality cooking implements, and cheaper procurement of ingredients. People eat at home because it's cheaper and they get exactly what they want. Neither of those need be exclusive to home meals.

[Gerry Burde @GerryBurde]

@wertform_mcm I bet it's 2x better

[Gerry Burde @GerryBurde]

@allTheYud Maybe even Late November without something groundbreaking

[Gerry Burde @GerryBurde]

But still $110 Billion of Office licenses are sold every year.

Microsoft 365 pulls in $110B+ annually with over 400M paid commercial seats. As a standalone company, that’s a ~$1.2T+ valuation.

Good luck getting corporate legal to redline contracts in Git, or enterprise auditors to debug a Python script instead of an Excel model. Office isn't just doc tools—it's the identity, compliance, and communication operating system of the Fortune 500. Can other solutions replace that stack? Probably.

[Gerry Burde @GerryBurde]

That's Wright’s Law (the Experience Curve)—and we’re watching electrochemistry run the exact same playbook solar perfected over the last 40 years. For every doubling of production we have a predictable learning rate that further reduces costs by 20-24%.

In photovoltaics, it’s known as Swanson’s Law: for every cumulative doubling of production volume, module costs decline by roughly 20–24%. That’s how solar modules crashed from over $100/W in the 1970s to under $0.10/W today. Lithium-ion—and stationary LFP in particular—runs on virtually the identical ~20% empirical learning rate.

The critical distinction is that Moore’s Law is governed by the calendar, while Wright’s Law is governed purely by cumulative manufacturing reps. It isn't just passive economies of scale; it’s continuous yield optimization. Just as solar squeezed out cost via diamond-wire wafer slicing, thinner wafers, and PERC-to-TOPCon efficiency jumps, battery gigafactories drove down costs by speeding up roll-to-roll slurry coating, moving to cell-to-pack (CTP) architectures, and crushing scrap rates from ~15% down to under 2%.

What makes this transition unprecedented is the compounding flywheel:

Ultra-cheap solar crashes midday wholesale power prices to zero (or negative).

That brutal price spread creates an irresistible arbitrage incentive for stationary storage.

Rapid BESS procurement accelerates global cumulative GWh capacity doublings.

Battery costs slide down the Wright’s Law curve, pushing tier-1 LFP cells to $30–$40/kWh and packs under $70/kWh.

We don’t just have one exponential clean energy technology—we have two separate ~20% learning curves directly feeding each other to permanently re-underwrite the economics of the grid.

[Gerry Burde @GerryBurde]

Not possible. The premise that one entity can reach recursive self-improvement (RSI) or artificial superintelligence (ASI) and swallow the entire global economy overnight overlooks a fundamental reality: intelligence is only an input, not the entire economic circuit.

Even if a lab develops a vastly superior model, several structural walls prevent a single company from capturing the entire market:

1. The Physical and Regulatory Latency of the Real World
Software scales at the speed of light; the physical economy scales at the speed of steel, concrete, and law.

Capital & Infrastructure: Deploying intelligence into the physical world requires energy grids, specialized fab capacity (semiconductors), hardware robotics, and supply chains. No single company can vertically own the global electrical grid, lithography machines, and industrial manufacturing simultaneously.

Sovereignty & Antitrust: No major nation-state—the US, the EU, or China—will permit a single private corporation to monopolize the entire cognitive infrastructure of society. Regulatory interventions, antitrust mandates, nationalization threats, and sovereign compute initiatives would fragment and counter a single monopoly long before it captures 100% of global GDP.

2. The Distribution, Integration, and Trust Moat
Replacing a pure software feature (like an MCP plugin or a wrapper app) is easy. Replacing deep enterprise and institutional integration is excruciatingly hard.

Institutional Inertia: Healthcare systems, financial institutions, defense departments, and legacy enterprises do not swap core systems overnight just because a smarter model exists. They operate on compliance, liability, audited processes, and organizational inertia.

The Liability Problem: Who goes to court when a system fails? A foundation model provider cannot legally underwrite every medical diagnosis, financial transaction, structural engineering calculation, and legal trial worldwide. Domain-specific intermediaries exist largely to absorb liability, integrate workflows, and ensure domain compliance.

3. Open-Source and Fast-Follower Dynamics
The technological frontier is never an isolated island.

Breakthroughs leak through academic papers, open weights, talent migration, and reverse-engineering of techniques (distillation, synthetic data, architectural tweaks).

As seen with open-source ecosystems (e.g., Llama, DeepSeek, Mistral), the performance gap between frontier closed labs and decentralized alternatives rarely remains wide enough for one company to permanently monopolize capability. An ecosystem of millions of developers, nation-states, and rival tech giants running open or decentralized models forms a counterweight that catches up rapidly.

4. Specialization Beats Centralized Monopoly (Comparative Advantage)
David Ricardo’s law of comparative advantage applies to intelligence just as it does to human labor.

Even if an ASI is theoretically superior at every task, compute is a finite resource governed by economic trade-offs (cost per inference, latency, and local context).

Smaller, specialized, on-device, or federated architectures will consistently be more cost-effective and private for discrete tasks than routing all global computation through a single centralized intelligence engine.

In Summary
A faster engine does not make the car company own the highways, gas stations, shipping fleets, and trucking logistics. It simply forces the entire surrounding system to adapt. The economy is not a monolithic program waiting to be rewritten by a single compiler; it is an organic, adversarial, highly regulated network of billions of nodes that will adapt, counter, and distribute intelligence rather than surrender to a single player.

[Gerry Burde @GerryBurde]

@siliconcodesign It's not really possible with current material science to overbuild AI compute capacity.

[Gerry Burde @GerryBurde]

@luke_veil @JStefanop1 It's kind of the basics... https://t.co/cHHRLwAA5I

[Gerry Burde @GerryBurde]

@nicbarkeragain Doing something similar with rust. Just need to get my agents on board with it, they still default to python no matter the instructions.

[Gerry Burde @GerryBurde]

@buildwithrajath Why not both? Did they offer both? 2x usage is clearly a better option. Most of us have other subscriptions to balance usage over so we don't exceed a 5 hour window.

[Gerry Burde @GerryBurde]

@thsottiaux It kind of depends on how much of the weekly the 5h limits give. If it's like something like 25% of a weekly in each 5hr bucket then yes.