If the Internet Collapsed "Distance," AI Collapses Time

If the core function of the internet is connectivity and the core value is the collapse of distance, then the core function of artificial intelligence is cognition and the core value is the collapse of time. 

If the internet makes physical location less important, AI makes complexity less important, reducing the time to derive insights. 

But both the internet and AI are going to disintermediate value chains, removing distribution functions and providers. 

Dimension	Internet	AI
Core Function	Connectivity: linking people, machines, data, and services across networks.	Cognition: performing tasks that require perception, reasoning, analysis, prediction, or decision support.
Primary Value Created	Eliminating distance: collapsing geography; enabling instantaneous communication and access.	Saving time: collapsing effort; automating, accelerating, or augmenting cognitive tasks.
Economic Logic	Reduces transaction and coordination costs associated with physical separation.	Reduces cognitive labor costs and enhances productivity by automating thinking tasks.
Primary Constraint	Bandwidth, latency, physical infrastructure (fiber, spectrum).	Quality of data, model capability, alignment with goals, compute.
Main Units of Scarcity	Transport capacity (Mbps/Gbps), access points (ports, routers), spectrum.	Compute, data quality, reasoning ability, task generalization.
User Experience Shift	From location-dependent to location-independent access.	From manual decision-making to automated or assisted decision-making.
Industrial Impact	Creates global digital markets; enables remote work, cloud services, platform economies.	Automates white-collar workflows; reshapes knowledge industries; introduces agentic systems.
Business Models	Subscription access, metered usage, advertising, platform marketplaces.	API usage, per-inference billing, embedded intelligence in existing software, agentic task fees.
Strategic Advantage	Owning the pipes, connectivity footprint, spectrum, and interconnection points.	Owning the models, data, workflows, and user attention for cognitive automation.
Regulatory Focus	Universal access, net neutrality, infrastructure competition.	Bias, transparency, safety, copyright, workforce displacement.
Transformation Pattern	Disintermediation of distance-dependent middlemen (retail → e-commerce; media → streaming).	Disintermediation of cognition-dependent middlemen (analysts, coordinators, support roles via agents).

So one way of understanding AI is to view it as a new form of infrastructure, as is the internet, as was electricity or railroads. In that view, potential over-investment happens because the new infrastructure has to be created, and not because of a mania or bubble over asset values that are illusory. 

That might temper some of the concern over AI asset valuations or investment magnitude, which can appear excessive in the near term, and might well be, in some instances. Such early over-investment tends to happen when a new general-purpose technology emerges, and especially when that GPT involves infrastructure.  

Historically, transformative infrastructure projects such as railroads experienced periods of perceived "over-investment," where excess capacity was common before widespread economic and societal adoption caught up. 

The U.S. railroad boom of the late 19th century and the electricity grid’s rollout involved capital surges, initial overbuilding, and even bankruptcies. However, over time, these investments generated foundational benefits, enabling entirely new industries and reshaping nations.

So although the superficial similarity between an irrational asset bubble and an infrastructure boom can exist, they are quite different. 

While a financial bubble features a disconnect between investment and credible returns, general-purpose infrastructure has long-term value, even if some amount of capital is misallocated. 

But that’s the issue right now: some see the infrastructure for a general-purpose technology being built; others see mostly speculation. It can be hard to tell the difference in the early going. 

Criterion	Productive Infrastructure	Speculative Excess
Cost-Benefit Analysis	Thorough, data-driven	Minimal or absent
Multiplier Effect	High, measurable output/wages	Weak, limited economic return
Demand Alignment	Supported by real user/market needs	Based on future hype, not evidence
Systemic Productivity	Positive spillovers	Neutral or negative impact
Asset Price Relationship	Aligned with long-term value	Driven by short-term speculation
Evaluation Rigor	Institutional, non-partisan reviews	Ad hoc, driven by momentum

Anthropic Expects 1 GW of Compute Power Online in 2026

Anthropic has announced a multibillion-dollar partnership with Google giving Anthropic access to one million Google Tensor Processing Units and more than one gigawatt of compute power by 2026.

It might not be unrealistic to expect the other leading model suppliers to keep pace, as the evolution of model generations has seemed to require an order of magnitude increase in power capability with each succeeding generation of models.

Facility	Organization	Power Capacity	GPU Count
xAI Colossus	xAI	~100-150 MW	200,000 H100s
Meta AI Research Cluster	Meta	~50-100 MW	100,000+ GPUs
Microsoft Azure AI	Microsoft	~200-300 MW	Distributed
Google TPU Clusters	Google	~150-250 MW	Millions of TPUs

AWS CEO Projections for Future Training
Generation	Power Requirement Per Model	Timeline
Current (Gen2-3)	50-200 MW	2023-2025
Next Gen (Gen3-4)	1-5 GW	2026-2028
Gen5+	5-10+ GW	2029+

Planned/Under Construction Mega Data Centers
Project	Organization	Total Power Capacity	Investment	Timeline
Stargate (Total)	OpenAI/Oracle/SoftBank	10 GW	$500B	2025-2029
Stargate Abilene	OpenAI/Oracle	1-2 GW	$10B+	2025-2026
Stargate Phase 2	OpenAI/Oracle	4.5 GW	$300B+	2026-2028
Stargate Phase 3	OpenAI/SoftBank	1.5 GW	$50B+	2026-2027
UAE Stargate	G42/OpenAI	5 GW	TBD	2026+
Meta Louisiana	Meta	2-4 GW	$10B+	2026-2028
Microsoft Wisconsin	Microsoft	3-5 GW	$15B+	2026-2029
CoreWeave Pennsylvania	CoreWeave	0.5-1 GW	$6B	2025-2027

Compute Capability of Current LLM Models
Model	Organization	Release Date	Training Compute (FLOPs)	Estimated Cost
GPT-4	OpenAI	March 2023	~2×10²⁵	~$100M
Gemini Ultra	Google	Dec 2023	~10²⁵	~$100M
Claude 4 (Opus/Sonnet)	Anthropic	May 2025	~10²⁵	~$100M
GPT-4o	OpenAI	2024	~10²⁵	~$100M+
GPT-4.5	OpenAI	2025	~10²⁵ - 10²⁶	~$100M+
Gemini 2.5 Pro	Google	March 2025	~10²⁵ - 10²⁶	~$100M+
Grok 3	xAI	Feb 2025	10²⁶	~$500M-$1B
Grok 4	xAI	July 2025	~10²⁶ - 10²⁷	~$1B
Llama 3	Meta	2024	~10²⁵	~$100M

At least up to this point, each generation requires approximately 10 times more compute than the previous generation to achieve significant capability improvements.

Expected Compute for Coming Models
Model	Organization	Expected Release	Training Compute (FLOPs)	Estimated Cost
GPT-5	OpenAI	Late 2025-2026	10²⁶ - 10²⁷	$1B+
Claude Next	Anthropic	2026	~10²⁵ - 10²⁶	$1B+
Gemini 3.0	Google	2026	10²⁶ - 10²⁷	$1B+
Llama 4	Meta	2025-2026	10²⁶	$500M-$1B