Where AWS Sees Value in the AI Stack, What it Means for Investors

Andy Jassy’s recent letter to shareholders provides a way of thinking about where artificial intelligence startups will be created; where functional objectives can lead to new company revenue streams and how the value chain will develop. 

Jassy talks about bottom, middle and application layers of AI. Using the software stack as an analogy, this corresponds to infra, middleware, app layers. 

The bottom layer includes both hardware and software: AI foundation models (generative AI, for example);  the computing infra required to train models and generate inferences and the software that makes it easier to build these models. 

Jassy points out that virtually all the leading models have been trained on Nvidia chips, but that customers “have asked us to push the envelope on price-performance for AI chips.”

So Amazon Web Services has built custom AI training chips (Trainium) and inference chips (Inferentia). Those chips are used by Anthropic, for example. Other users include include Airbnb, Hugging Face, Qualtrics, Ricoh and Snap.

Customers building their own models must organize and fine-tune data, build scalable and efficient training infrastructure, and then deploy models at scale in a low latency, cost-efficient manner.

Amazon SageMaker is a managed, end-to-end service for preparing their data, managing experiments, training models faster, lowering inference latency and improving developer productivity, Jassy says. 

At a broader level, all that implies opportunities for rival graphics processor units, acceleration chips, generative AI models and AI “as a service” businesses, 

The middle layer is for customers seeking to leverage an existing model, customize it with their own data, and leverage a leading cloud provider’s security and features to build a GenAI application as a managed service, Jassy says. 

We might also liken the process of “rendering” to the middle layer as well. In computer graphics, rendering is the creation of 2D images from 3D models. In audio production, rendering refers to the process of creating a complete audio file from multiple different tracks. 

In video production, rendering (editing, formatting) might refer to the processes of creating a final version of the product, adding visual effects or formatting for specific delivery formats (resolution, frame rate). 

Amazon Bedrock is an example of this layer, including “Guardrails: to safeguard what questions applications will answer), “Knowledge Bases,” as well as “Agents” to complete multi-step tasks) and “Fine-Tuning” to keep teaching and refining models.

Customers using Bedrock include ADP, Amdocs, Bridgewater Associates, Broadridge, Clariant, Dana-Farber Cancer Institute, Delta Air Lines, Druva, Genesys, Genomics England, GoDaddy, Intuit, KT, Lonely Planet, LexisNexis, Netsmart, Perplexity AI, Pfizer, PGA TOUR, Ricoh, Rocket Companies, and Siemens, Jassy says. 

AI “as a service” presents one set of opportunities, but every commercially-viable AI model will require this sort of middle-layer support as well, often sourced from third parties. 

The top layer of the AI stack is the application layer. 

For Amazon that includes shopping assistants, smarter versions of Alexa, advertising, customer service and seller services, as well as coding support apps to write, debug, test and implement code. Such apps might also support moving platforms from older to newer versions, conducting queries across multiple data repositories, summarizing data, conducting conversations and taking actions as assistants.

We should already see investment shifting from generative AI models to applications (ways to use the models to solve business problems or conduct consumer operations and tasks. Some examples include apps aimed at industry verticals, horizontal functions such as customer service or coding, fraud detection, healthcare diagnostics or supply chain optimization. 

As always is the case for general-purpose technologies, early investment goes into creating infrastructure. Later investment broadens to create applications and use cases across multiple industries and functions. 

Startups will be the field for private equity firms, institutional investors and accredited investors. Most of the opportunities for consumer investors will come in the form of publicly-traded firms with some plausible involvement in bottom, middle and app layers (infra, middleware, end user and retail supplier use cases). 

Unless you own or work for a venture capital firm, or are an accredited investor, you will not be able to invest in startups oriented around artificial intelligence, leaving you with the task of identifying existing public firms that have some plausible direct relationship to AI. 

Eventually, as has proven to be true for the internet, most firms will have some indirect relationship to the internet, but that is not so helpful in identifying candidates for investment right now. 

As is true for just about any general-purpose technology before it (steam power, railroads, the internet, electricity) and other platforms that might not always be considered GPTs (highway systems, passenger air travel, mobile communications, telephone systems), infrastructure is where investments must be made first, before the full range of use cases develops. 

So for “regular people” the domain initially will be public firms with AI infrastructure operations: the compute power to run AI software, the products required to build AI models, make inferences, create applications or supply platforms and devices to run the models and make inferences, sustain the connectivity to processing nodes, create and run the data centers, provide AI functions as a service. 

Most of these infra segments includes firms one might own for other reasons as well (dividends, revenue growth, capital appreciation, real estate investments, software or information technology, content assets, connectivity). 

So, perhaps oddly enough, “AI investments” are pretty much the same as would be expected if one were instead searching for “digital economy” or “internet” investments, with perhaps a stronger weighting towards “picks and shovels” that create or sustain the infrastructure to run apps and provide experiences. 

Analyst Cody Willard suggests an 11-layer model focused on AI infrastructure, including some private firms or open-source initiatives, but also focusing on public firms plausibly involved in creating AI infrastructure, applications and content. 

Chips, servers, data centers, cloud computing, data management, algorithms, models, internet connectivity and end user devices are perhaps the clearest examples of AI infra. But some might also include content, enterprise AI-enabled applications or advertising as layers of the AI value chain. 

source: Cody Willard, MarketWatch 

Looking at infra as 11 or more layers, that might suggest a layer one (chips) including

• Silicon (Nvidia, duh!), Intel, AMD, Qualcomm and Broadcom (add Microsoft, Alphabet, Meta and Apple to the extent they are developing their own AI chips as well)

• Silicon design services such as Cadence Design Systems, Synopsis  or Autodesk

• Application Specific Integrated Circuit (ASIC) designer such as Broadcom , Marvel, Intel, Advanced Micro Devices and Qualcomm

• Silicon intellectual property including Arm, Intel

• Semiconductor equipment such as ASML, Advanced Materials, Lam Research, KLA Corp. and Teradyne

• Foundries including TSMC, Intel, Samsung, Global Foundries

• Memory (SK Hynix, Samsung, Micron, Western Digital, Seagate

• Machine Learning languages including PyTorch (open-sourced from Meta), TensorFlow (open-sourced from Google), Keras, Microsoft Cognitive Toolkit, Theano, Apache MXNet, Chainer, JAX, TensorFlow.js

Layer two might focus on servers, including:

• Server design (Dell, Hewlett-Packard Enterprises, Super Micro, IBM, Lenovo, Cisco, Oracle, Fujitsu, Quanta Cloud Technology, Inspur

• Server manufacturing including Foxconn, Flex, Jabil, Sanmina Corp., Pegatron Corp., Celestica, Wistron Corp., , Quanta Computer,, Compal Electronics, Inventec Corp..

• Distribution partners include Ingram Micro, Aero Electronics and CDW.

• Server installation services (IBM, Schneider Electric, Vertiv Holdings Co., Hewlett Packard Enterprise, Super Micro, Dell

Layer three can be viewed as data centers:

• Data center design and construction (Holder Construction, Turner Construction, Jacobs, Fluor Corporation, AECOM, Syska Hennessy Group, Corgan, Gensler, HDR, Mastek, Dycom

• Data center Infra, especially cooling (Schneider Electric, Johnson Controls, Carrier Global Corporation, Honeywell International Inc., Siemens AG, Super Micro, Dell, Hewlett-Packard Enterprises

• Electric components, including renewable energy, including Enphase Energy, Inc., Solaredge, First Solar, Tesla, Inc., SunPower Corporation, Schneider Electric, ABB Ltd., Eaton Corporation

• Electrical power suppliers (PNM Resources, NextEra Energy, Duke Energy Corporation, Dominion Energy), Southern Company, Exelon Corporation, American Electric Power Company, PG&E Corp., Consolidated Edison, Xcel Energy, Entergy Corp.

• Electric utility infra (General Electric, Siemens Energy, Mitsubishi Heavy Industries, Toshiba Corporation, Hitachi Ltd., ABB Ltd., BWX Technologies, Doosan Heavy Industries & Construction

• Raw materials such as copper, gold, plastic (oil) and silver (Freeport-McMoran, Newmont Corp., Barrick Gold Corp., Franco-Nevada Corp., Freeport-McMoRan Inc., Southern Copper Corporation, BHP Group, ExxonMobil, Chevron, ConocoPhillips

• Networking and interconnect gear (Cisco Systems, Arista Networks, Inc., Juniper Networks, Inc., Broadcom Inc., NVIDIA Corp.), F5 Networks, Extreme Networks, Inc., Dell Technologies, Marvel, Applied Optoelectronics, Viavi Solutions, MaxLinear, Emcore, Nlight

Layer four might be envisioned as the “cloud” layer, including: 

• Cloud data centers (Amazon Web Services, Microsoft Azure, Google Cloud Platform, Alibaba, Oracle, Tesla, Meta

• Data center real estate investment trusts (Equinix, Digital Realty Trust, Inc., CyrusOne Inc. (KKR), CoreSite Realty Corporation (subsidiary of American Tower), QTS Realty Trust (Blackstone)), Iron Mountain Inc., DigitalBridge

• Cloud computing as a service providers (AWS, Microsoft Azure, GCP, Oracle, Meta, Tesla, Alibaba

• Inference As A Service (NVIDIA, Amazon Web Services, Google Cloud AI Platform, Microsoft, Cloudflare, Akamai

Layer five might be viewed as the data layer, including functions such as data gathering and input, machine vision, data sources:

• Machine vision including Tesla, Rockwell Automation, Zebra, Cognex Corp., Keyence Corp., OMRON Corp., Basler AG, Teledyne Technologies, ISRA VISION AG

• Consumer data sources (shopping, other behavior including Meta, ByteDance, Google, Apple, Amazon, Snap, Pinterest, Yelp, Tencent, Reddit, Etsy,, Wayfair, Walmart

• Financial data (Apple Pay, Google Pay, JPMorgan Chase & Co., Visa Inc., Mastercard Inc., Discover Financial Services, PayPal Holdings, Inc., Square, Inc., Robinhood Markets, Inc., Morgan Stanley, The Charles Schwab Corp., Fair Isaac Corp., TransUnion, Equifax Inc.

• Location, travel data (Apple Inc., Alphabet Inc., Verizon Communications Inc., AT&T Inc., T-Mobile US, Inc., Uber Technologies, Inc., Lyft, Inc., Expedia, Tripadvisor

• Enterprise Data (ServiceNow, Apple, Salesforce.com, Inc., Oracle Corp., Microsoft Corp., Alphabet Inc., Dropbox, Inc., Box, Inc., Workday, Inc., AutoDesk, Adobe, Dassault Systèmes, PTC Inc., Ansys, Inc., Trimble Inc., Siemens AG, AVEVA Group plc, Bentley Systems, Inc..

• Content (Disney, Sony, Spotify, Netflix, Warner Brothers, Paramount, New York Times, Fox, Simon & Shuster, Random House

• Data Management (Amazon Web Services, Alphabet Inc., Microsoft Corp., Snowflake Inc, Datadog, MongoDB, Inc., Oracle Corp., Confluent, Inc., Broadcom Inc., Alteryx, Inc., International Business Machines Corp., Cisco Systems, Inc. 

Layer six is the algorithm and model layer:

• Algorithms (OpenAI, Google Deepmind, Tesla AI, Meta Labs

• Large language models (OpenAI, Google, Microsoft, Anthropic, Perplexity) and training

• Libraries: Hugging Face

Layer seven is the applications layer:

• Generative AI as an app (ChatGPT, Gemini, Anthropic, xAI Grok, LLaMA, Stability AI, Mistral, Mosaic, Amazon

• Apps using LLM (Microsoft Copilot & Github, Office365, Google Workspace, Duolingo, Alexa, Siri, Spotify, Palantir, Autodesk, Unity

• Data analysis (Snowflake, Oracle, Datadog, AWS, Google, Azure

• Content Creation (The Walt Disney Company, Netflix, Inc., Electronic Arts Inc., Warner Bros. Discovery, Inc., Paramount Global, Sony Group Corp. , Comcast Corp. , Activision Blizzard, Inc. (Microsoft), Electronic Arts, Take-Two Interactive Software, Inc., Spotify Technology S.A., Lions Gate Entertainment Corp, Fox

• Cybersecurity (Palo Alto Networks, Fortinet, , CrowdStrike Holdings,, Zscaler, , Check Point Software Technologies Ltd. , CyberArk Software Ltd., Okta, Inc., FireEye, Inc.

Layer eight might be edge networking:

• Hardware and servers (Intel, AMD, NVIDIA, Dell, HPE)

• Content Delivery Networks (Cloudflare, Akamai, Fastly, AWS, GCP, Azure)

Layer nine might be advertising:

• Venues (Meta, Google, Amazon, Snapchat, Pinterest)

• Ad placement and services (The Trade Desk, Unity, Applovin)

Layer 10 might be networking:

• Tower infra (Crown Castle International Corp., American Tower, SBAC)

• Access providers (Starlink, Verizon Communications, AT&T, Lumen Technologies, Charter Communications, T-Mobile, Comcast, Comtech, ViaSat, Iridium Communications,  Gogo

Layer 11 includes end user devices:

• Consumer devices including Apple iPhone, iPad, Mac, PCs, Google Pixel phones, Meta Glasses, Lenovo PCs

• Enterprise devices such as robots (Tesla Optimus, Mitsubishi, Kawasaki, Epson, Universal Robots, Omron, Yaskawa Electric, Fanuc, Kuka, Denso, ABB

• AI machinery (tractors, cranes, containers, and boats made by Caterpillar, John Deere, Trimble)

• AI satellites and spacecraft: SpaceX, Rocket Lab, Intuitive Machines, Optimus by Tesla)

• AI Drones by AeroVironment, Lockheed Martin, Boeing, Northrop Grumman

Some of us might argue that mass market investors should view virtually all these assets as categories we’d consider for other reasons, despite their AI potential, as it might be some time before AI revenues are material. 

In Willard’s 11-layer model, some of us might consider much of layer seven and virtually all of layers eight through 11 as being part of the broader computing and internet value chains, and not specifically powered by AI potential. 

And parts of layers one through six would be required to support modern computing, even if AI did not exist. The point is that the AI value chain overlaps substantially with the internet value chain. With a few specific exceptions, Nvidia being the primary example, virtually all the other assets in the developing AI value chain would also be candidates for ownership as part of the internet value chain.

Private Equity, Overbuilder and Telco FTTH Payback Models are Very Different

Firms backed by private equity have different business models than other long-term operators of connectivity assets. PE-backed firms aim to create value (typically double the asset value within seven years) and then sell the assets. 

That is a different model than used by connectivity service providers who operate for the long term, where fundamental issues of free cash flow, revenue growth and profit, as well as the ability to pay dividends, are the key constraints. 

And so it is with investors in fiber-to-home assets. 

Back in the heady days of 1996, when the Telecommunications Act of 1996 became law, business models for firms providing connectivity services changed in a big way. For legacy providers, maintaining market share became the key issue. For attackers, gaining share became the obvious key issue. 

Beyond that, the imperatives were different. Legacy providers, operating their businesses for the long haul, could not adopt the “fast growth rather than profits” models as used by many attackers. At a time of “easy money” and “we want you to grow fast” attitudes of key investors, that made sense for attackers.

And, as has been true for many software startups, long-terms operating profits were not the goal. Instead, fast growth in a “hot” area was the objective, since such firms had reasonable expectations they would simply be bought out at some point before they ever reached “terminal value.”

That, at least, is what one has to assume when looking at the costs of FTTH networks and costs to actually connect customers and earn a profit on those services.

The reported cost per-home-passed (CPHP) for underground FTTH deployments ranged from $1,600 to $2,600, according to a recent estimate by Cartesian researchers. The CPHP for aerial deployments was lower than those of underground, ranging from under $700 to $1,500 for respondents in suburban and urban environments, and $1,300 to $2,700 in more rural areas. 

source: Fiber Broadband Association 

Actually connecting a paying customer adds another $600 to $830 in drop costs. 

source: Fiber Broadband Association 

So the per-home cost of serving a paying customer includes an attributed cost of building the network; an assumption about take rates and then the cost of the drop and installation; plus operating and marketing costs. 

Take rates matter. At a 50-percent take rate, for example, the per-customer cost of the network can range from $2,600 to perhaps $5,200, with an additional $600 to $800 in drop costs, for a per-customer network cost ranging from a “best case” of perhaps $3,200 up to perhaps $6,000. 

But that is just the network platform. One would have to add in operating and marketing costs, plus any debt service and loan principal repayments. Operating and marketing costs might range from about $210 per year to $800 per year, per customer, according to some estimates. 

Cost Category	Low Estimate ($/year/subscriber)	High Estimate (/year/subscriber)	Sources
Network Infrastructure	$100	$500	FTTH Council: $200-$300,  Deloitte: $300-$500
Operations & Maintenance (O&M)	$25	$75	FTTH Council: $40-$60. Analysys Mason: $25-$35
Customer Acquisition (CAC)	$50	$150	BroadbandNow: $50-$100, Analysys Mason: $60-$150
Customer Care & Billing	$25	$50	Analysys Mason: $25-$35,  Leichtman Research Group: $30-$40
Marketing & Sales	$10	$30	Analysys Mason: $10-$20,  Leichtman Research Group: $15-$25
Total Operating Cost	$210	$805	Sum of individual ranges

And one might have to add interest charges and eventual debt principal repayment in addition to those charges. 

And there is a possible additional range of investments as well. Some firms must first acquire copper-based legacy telco assets first, before starting the FTTH upgrade, either to own and operate over the long term, or to sell the assets in five to seven years. 

Transaction	Date	Buyer	Seller	Asset Type	Homes Passed (M)	Price (USD Billion)	Cost per Passing (USD)	Source
Brightspeed - Lumen assets (20 states)	Oct 2022	Brightspeed	Lumen	Fiber	0.3	3.0	10,000	Reuters
Consolidated Communications - NewWave Communications	Aug 2022	Consolidated	NewWave	Fiber	0.18	0.65	3,611	Fierce Telecom
Windstream - MetroNet Holdings (FL)	Aug 2022	Windstream	MetroNet	Fiber	0.06	0.28	4,667	Fierce Telecom
Frontier Communications - Verizon (WA, OR)	Dec 2021	Frontier	Verizon	Mixed (Fiber & Copper)	0.14	1.05	7,500	Fierce Telecom
Allo Communications - Lincoln Telephone & Telegraph	Nov 2021	Allo	Lincoln	Mixed (Fiber & Copper)	0.11	0.21	1,909	TelecomTV
Ziply Fiber - US Cellular assets (WA, OR)	Oct 2021	Ziply	US Cellular	Fiber	0.12	0.51	4,250	Fierce Telecom
CNSL - Searchlight Investment	Jan 2020	Searchlight	CNSL	Mixed (Fiber & Copper)	0.71	0.425	600	CNBC

In many cases, the capital investment to acquire assets is equal to, or more than, the cost to add the FTTH upgrade. But that’s where the business case lies. If one assumes a copper asset can be purchased for $600 to $800 per passing, but then an upgraded FTTH asset can be sold for $5,000 to $10,000 per passing, that is the business case for making all the investments in FTTH. 

It might still be a difficult business case for a shorter-term owner, but “buying copper assets; upgrading to FTTH and then selling” can work. 

The payback for longer-term operators always has been equally challenging, if not more challenging, and has gotten arguably tougher as total account revenues including voice and video entertainment have dwindled, forcing the payback model to be based on home broadband alone. 

The main point is that FTTH payback models for private equity investors and service providers are quite distinct. What makes sense for a PE firm might not always make sense for a legacy fixed network service provider or an “overbuilder.” 

That is perhaps one reason why GFiber (owned by Alphabet) has not purchased copper telco fixed network assets before upgrading them. As with other “overbuilders,” GFiber has simply built its own greenfield FTTH networks from scratch.