Amazon has at Least 160 AI Initiatives Underway, Sources Say

Amazon's retail business supports more than 160 AI initiatives, including the Rufus shopping assistant and Theia product-image generator, according to a report by Business Insider. 

Other AI projects in the works include:

• A vision-assisted package retrieval service that uses computer-vision technology to help drivers quickly identify and pick the correct packages from vans at delivery stops.

• A service that automatically pulls in data from external websites to create consistent product information.

• A new AI model that optimizes driver routing and package handling to reduce delivery times and improve efficiency.

• An improved customer service agent that uses natural language to address customer return inquiries.

• A service that automates seller fraud investigations and verifies document compliance.

Last year, Amazon estimated that AI investments by its retail business indirectly contributed $2.5 billion in operating profits. Those investments also resulted in about $670 million in variable cost savings, Amazon sources said.

Turns Out, Voice is a Product Like Any Other

For many of us who were familiar with the voice business, it seemed inconceivable that fixed network voice calling would be a product like any other, with a life cycle of growth, maturation and then decline. But that has happened. 

Surveys of U.S. consumer use of landline telephone networks  (2022 to 2024) suggest the percentage of U.S. households or adults with any type of landline phone service ranges between 24 percent and 30 percent, down from 90-percent or higher levels  in the early 2000s.

One reason is that about 70 percent or more of U.S. consumers rely on mobile phones for voice communications. 

U.S. carrier of last resort laws impose specific obligations on incumbent phone companies to ensure basic telephone service is available to all residents. The core principle is the obligation to provide service to any customer, irrespective of how difficult or costly it might be to reach them.

Such laws have cost implications for telcos trying to modernize their networks, as a carrier of last resort  cannot simply abandon service or cease operations in its designated territory without obtaining permission from state public utility commissions. 

As a practical matter, that includes the copper access networks few customers use anymore. The obligation is understood to mean that "basic service" (voice connections over copper networks) must be maintained. 

The key issue is that service providers would vastly prefer to retire the lightly-used copper networks and replace them with modern fiber optic networks, for example.

"NeoCloud" Emerges

The term "neocloud" is used to describe a new generation of cloud infrastructure companies such as CoreWeave, Lambda Labs, Crusoe and others that are distinct from the traditional hyperscalers like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud Platform (GCP)."Neocloud" (literally 

The neocloud providers specialize in AI infrastructure, especially compute power specifically tailored for artificial intelligence and machine learning workloads, especially using graphics processing units. 

Some would argue that neocloud providers offer computing that is more flexible and customized than tends to be offered by hyperscalers, offering custom infrastructure setups, shorter contract commitments, or usage-based pricing models.

Some providers, including CoreWeave and Nebius Group,  design their entire stack around AI, with support for AI model training, fine-tuning, and inference workloads. 

The Tomb is Empty

 

Happy Easter, brothers and sisters.

How LLMs Work, by Andrej Karpathy

Andrej Karpathy, Eureka Labs founder and computer scientist (Tesla, OpenAI), explains how language models work, and are built. You'll need about 3.5 hours to view the whole video, but it covers transformer networks, training (human and algorithmic; labeling) and reinforcement learning. 

Google Loses Antitrust Lawsuit, Not a Good Precedent for Meta

Google has been found guilty under the Sherman Act United States v. Google LLC 

Of illegally monopolizing advertising markets for publisher ad servers and ad exchanges used in open-web display advertising.   

The ruling is that Google engaged in a series of anti-competitive acts to willfully acquire and maintain monopoly power in the ad technology market, using acquisitions of ad software companies (such as DoubleClick and AdMeld) to unlawfully shore up market share and stifle competition.

The ruling probably does not bode well for Meta, which faces similar antitrust action brought by the Federal Trade Commission. Principally, the FTC alleges that Meta illegally maintained a monopoly in the "personal social networking services" market through anticompetitive conduct. 

The core of the complaint focuses on the acquisition of competitors, as was the finding in the Google ad tech antitrust decision. 

The FTC argues Meta employed a "buy-or-bury" strategy, specifically targeting its acquisitions of Instagram in 2012 and WhatsApp in 2014 to neutralize potential competitive threats and maintain its dominance.

What matters are the remedies the government entities might propose, assuming Meta likewise is found guilty of monopolistic behavior. 

For Amazon, AI is "Existential"

Though observers remain concerned about the huge amounts of capital being invested in artificial intelligence by hyperscalers, they arguably are doing so because AI poses disruptive threats comparable to past general-purpose technologies. It is, in other words, quite literally "existential," the very existence of a thing.

In other words, as happens with GPTs, AI could  fundamentally change how value is created and captured across industries such as search and e-commerce. Indeed, that happens frequently when a GPT arises. 

Electricity enabled entirely new industries built around electric motors (factories, appliances) and communications (radio, eventually TV). 

The internet revolutionized communication (email, social media), information access (web search), commerce (e-commerce), and entertainment (streaming). It decimated industries like print media and physical retail while creating whole new replacements led by different firms than led the legacy industries. 

Looking only at e-commerce, there already are glimmers of potential change. According to an analysis by Adobe Analytics, AI-driven traffic to retail websites surged 1,200 percent during the most-recent 2024 holiday shopping season.

source: Adobe 

Adobe’s survey of 5,000 U.S. consumers suggests 39 percent have used generative AI for online shopping, with 53 percent planning to do so in 2025.

Consumers use generative AI for research (55 percent of respondents), receiving product recommendations (47 percent), seeking deals (43 percent), getting present ideas (35 percent), finding unique products (35 percent) and creating shopping lists (33 percent).

According to a survey conducted by Adobe Analytics based on one trillion visits to U.S. retail sites, the trend extended beyond the holiday season as traffic to retail sites from AI-driven searches increased 1,200 percent in February compared to July 2024 and has doubled every two months since September 2024.

On Cyber Monday alone, traffic from generative AI sources soared by 1,950 percent from last year’s event, Adobe says. 

Compared to consumers coming from non-AI traffic sources (including paid search, affiliates and partners, email, organic search and social media), consumers coming from generative AI sources show eight percent higher engagement as they dwell sites for a longer period of time. 

These visitors also browse 12 percent more pages per visit, with a 23 percent lower bounce rate, Adobe notes.

source: Adobe 

At least for the moment, conversion (visits that become purchases) is the one area where AI lags other sources of traffic. 

Traffic from generative AI sources is nine percent less likely to convert compared to other sources of traffic. But conversion rates are improving. In July 2024, the AI conversion gap was 43 percent.

source: Adobe 

As you might therefore guess, this poses threats to Amazon. 

Generative AI-powered Search engines and chat assistants are becoming increasingly important sources of consumer traffic for online web stores. 

In January 2025 Amazon had about 455 million monthly unique visitors. But AI assistants and chatbots collectively are getting three-digit millions of daily interactions. 

source: Similarweb 

And ChatGPT already is moving in the direction of adding shopping and e-commerce features. 

The point is that GPTs tend to disrupt industries and economies. So it makes sense that Amazon would invest heavily in AI to protect its commerce business, as Google will invest to protect its search business. 

And sometimes that works. Manufacturing firms previously reliant on steam, water, or manual power redesigned their factories around electric motors.

Energy companies focused initially on coal gas for lighting and kerosene for lamps pivoted their refining processes to prioritize gasoline production as demand shifted from lighting oil to fuel for internal combustion engines.

Likewise, IBM, originally a leader in tabulating machines and typewriters,  transitioned into the mainframe computer era. Walmart and Target added significant online commerce to their place-based operations. 

Many legacy content firms and industry segments were battered by the internet’s alternatives, but a few have managed to stabilize and even grow their online businesses. 

And financial firms have moved rapidly to embrace online commerce as well. The point is that the legacy providers are not without weapons in the fight to retain their relevance and even dominance. 

Still, the hyperscalers are investing so heavily in AI for obvious reasons: AI poses a genuine threat to their core business models.

Language Model Progress Blows Away Moore's Law

Language models are improving at a blistering pace, far outstripping what we have come to expect from computing in general and Moore’s Law in particular. Where Moore’s Law has suggested chip density doubles about every 18 months or so, AI language models have been improving nearly 300 times faster. 

The cost of querying an artificial intelligence model that scores the equivalent of GPT-3.5 (64.8) on MMLU, a popular benchmark for assessing language model performance, dropped from $20.00 per million tokens in November 2022 to just $0.07 per million tokens by October 2024 (Gemini-1.5-Flash-8B), a more than 280-fold reduction in approximately 18 months, according to Stanford University’s Human Centered AI Institute. 

source: Stanford University HAI 

Depending on the task, LLM inference prices have fallen anywhere from nine to 900 times per year.

At the hardware level, costs have declined by 30 percent annually, while energy efficiency has improved by 40 percent each year, HAI’s 2025 AI Index Report says.

How Big is GPU As a Service Market?

Opinions seem to differ on the importance of language models and therefore  graphics processor unit operations provided “as a service” by cloud computing as a service giants including Amazon Web Services, Google Cloud and Azure, for example. 

Current forecasts suggest that the single-digit billions (U.S. dollars) GPU as a service market could grow to about $36 billion in annual revenues within five years. While significant, such revenue forecasts might not be so large in relation to the capital investment being made to support GPU as a service capabilities. 

GPU as a Service Market Estimates  (2025-2030)
Year	Global Market Size (USD Billions)	U.S. Market Size (USD Billions)	Global Growth Rate (%)	U.S. Growth Rate (%)	Sources
2025	$8.20	$3.40	31.50%	29.80%	Grand View Research, MarketsandMarkets
2026	$11.10	$4.50	35.40%	32.40%	Gartner, IDC
2027	$15.20	$6.10	36.90%	35.60%	Forrester Research, Allied Market Research
2028	$20.70	$8.30	36.20%	36.10%	Mordor Intelligence, Technavio
2029	$27.50	$10.90	32.90%	31.30%	Emergen Research, IMARC Group
2030	$35.80	$14.10	30.20%	29.40%	Fortune Business Insights, Research and Markets

But GPU as a service is a subset of “AI as a service,” which is generally an order of magnitude greater than GPU as a service in any given year. 

AI as a Service Market Revenue Estimates (2025-2030)
Year	Global Revenue (USD Billions)	U.S. Revenue (USD Billions)	Global Growth Rate (%)	U.S. Growth Rate (%)	Sources
2025	$26.50	$11.80	34.20%	32.60%	Based on pre-2025 trends from Markets and Markets, Grand View Research
2026	$36.80	$16.10	38.90%	36.40%	Projections extrapolated from Gartne³, IDC forecast models
2027	$52.10	$22.40	41.60%	39.10%	Aligned with long-term trends identified by Allied Market Research
2028	$73.50	$31.20	41.10%	39.30%	Derived from Mordor Intelligence⁶, Technavio market analyses
2029	$101.90	$42.80	38.60%	37.20%	Based on growth curves projected by Emergen Research
2030	$138.60	$57.90	36.00%	35.30%	Consistent with Fortune Business Insights, Research and Markets

It also is fair to note that observers disagree a bit about which industry verticals will be the biggest users and beneficiaries of GPU as a service, as is the case also for estimates of which industries are likely to be the greatest beneficiaries and users of AI as a service. 

There seems general agreement that some industries will probably not be big users of either AI or GPU as a service. Agriculture, construction, hospitality and tourism, mining and government tend to be among those industries. 

Industry Vertical	Primary Operations	Why Unlikely to Be High GPUaaS Users	Key Constraints
Agriculture,  Forestry	Crop production, livestock management, forestry operations	Limited use of LLMs or AI; focus on IoT and basic analytics rather than compute-intensive tasks	Low data complexity, cost sensitivity, and limited scalability of AI applications
Construction	Building infrastructure, urban development, heavy machinery operations	Minimal reliance on LLMs; AI use limited to project management and basic design, not GPU-intensive	High capital costs prioritize physical assets over advanced compute infrastructure
Hospitality, Tourism	Hotels, restaurants, travel services, event management	Basic AI for customer service (e.g., chatbots) doesn’t require extensive GPU resources; low data volume	Focus on human-centric services, limited need for real-time complex processing
Public Administration	Government services, regulatory compliance, public policy implementation	Constrained by budgets and data privacy; AI use limited to basic automation, not large-scale LLM training	Bureaucratic inertia, regulatory restrictions, and preference for on-premises systems
Education (Traditional)	K-12 schools, universities, vocational training	Limited use of LLMs beyond administrative automation; cost barriers and focus on human-led instruction	Budget constraints, low computational needs, and ethical concerns around AI use
Mining, Quarrying	Mineral extraction, resource exploration, heavy equipment operations	AI limited to predictive maintenance and geospatial analysis, not requiring high GPU compute	Harsh environments, focus on physical operations, and low data-driven workflows
Waste Management,  Remediation	Waste collection, recycling, environmental cleanup	Minimal AI adoption; basic analytics for logistics don’t demand GPUaaS	Low-margin industry, limited scalability of AI, and focus on operational efficiency

But some industry verticals are generally considered to be heavier users. Drug discovery in the pharmaceutical industry; fraud detection and risk modeling in financial services and code generation in the technology business provide examples. 

Autonomous vehicles and content creation are other verticals where heavy use of AI is likely or necessary. 

Industry Vertical	Primary GPUaaS Use Cases	Key Drivers for Adoption	Estimated Impact
Healthcare, Life Sciences	Drug discovery, genomic analysis, medical imaging, virtual health assistants	Faster drug development, improved diagnostics, personalized medicine	High: Accelerates life-saving innovations and reduces R&D costs
Financial Services	Fraud detection, risk modeling, algorithmic trading, customer service automation	Low-latency predictions, regulatory compliance, competitive edge	High: Enhances accuracy and speed in high-stakes transactions
Technology,  Cloud Services	NLP, code generation, enterprise AI platforms, cloud-based AI services	Scalable AI infrastructure, market leadership in AI services	Very High: Powers AI ecosystems and serves other industries
Media,  Entertainment	Content creation, video transcoding, digital avatars, gaming AI	Demand for immersive, personalized content and real-time processing	Moderate-High: Drives innovation in creative and gaming sectors
Automotive,  Transportation	Autonomous driving, traffic optimization, vehicle design simulation	Safety, scalability of autonomous systems, synthetic data generation	High: Critical for advancing self-driving technology
Retail,  E-commerce	Chatbots, personalized marketing, inventory management, sentiment analysis	Enhanced customer experience, operational efficiency	Moderate-High: Boosts sales and customer satisfaction
Telecom	Network optimization, virtual assistants, predictive maintenance	Cost reduction, improved service quality, data-heavy operations	Moderate: Enhances efficiency in a competitive, data-driven industry