btov Partners
Published in

btov Partners

Finding Gold in the AI Value Stack

Where are the best entrepreneurial opportunities in the mainstream adoption of machine learning?

A consensus is emerging in tech: The next big thing is going to be AI, or more specifically, the broad adoption of machine learning (ML). Now that the S-curve of innovation in mobile is clearly flattening out (cf. Apple’s latest sales results), the tech industry needs another shiny object, and it’s increasingly clear that ML is it. Obviously there’s a lot of excitement around ML, but also a lot of fear and doubt, along with pure confusion about where the market is going. Why is that?

It’s a question of complexity. Huge technology waves are never monolithic. They always consist of a whole ecosystem of components that work together in creating value for end users, building on each other to provide a complete result — something you could call a value stack. AI’s value stack is very complex, but that’s not a huge difference from past technology waves.

The Last Wave: Mobile’s Value Stack

We intuitively might think of the smartphone revolution as something that is quite simple to understand and visualize: Apple’s iPhone seems to symbolize it perfectly. It’s an iconic product that defined an entire decade-long wave of innovation, right?

But it’s actually not nearly that simple. Think about all the other components that were necessary to make the iPhone (and its Android cousins) so successful:

  • Mobile networks and all the complexity (and gigantic investment) behind them
  • New mobile chip technology, particularly ARM-based processors
  • New screen technology, particularly multi touch-capable touchscreens
  • Mobile operating systems and related app stores (both owned by Apple for the iOS platform, but not by Android hardware makers)
  • Development tools and hosted backend systems that make developers more productive
  • Countless apps produced by independent software makers
  • Mobile marketing firms that enable app install campaigns
  • System integrators and mobile app developers that build apps for businesses
The Mobile Value Stack

Fortunes were (and are) made on all of these layers. The iPhone was just the most visible incarnation of this entire ecosystem of innovation that comprises tens of thousands of companies.

Many of the most iconic companies of the smartphone age used this whole value stack to create something unique — by applying all this power to a so far undisrupted domain. Uber is the most obvious example. You might not think of Uber as a “smartphone startup”, but without the entire value stack of mobile, this new type of transportation would have never been possible. There are of course countless other examples: Spotify, Instagram, Whatsapp, WeChat, Waze, Twitter, Snapchat, Tinder, payment apps like Square, SumUp or Venmo, the whole podcast industry, and so on.

Is the AI Value Stack Similarly Complex and Promising?

When people currently discuss AI, most attention is captured by specific, fairly narrow applications of deep learning neural networks: Google’s DeepMind team building the best Go and chess-playing algorithms; ML filters that beat the best doctors in analyzing MRI scans; facial recognition that performs better than humans; voice assistants like Alexa and Google Assistant; and, of course, self-driving cars.

All this attention is not surprising for three reasons:

  1. The field of deep learning has seen the most remarkable breakthroughs in the past few years, and there’s rarely a week when not another amazing achievement is published.
  2. Deep learning networks are opaque — we don’t understand in detail why exactly a particular model works that well, and that lends them a sense of mystery, bordering on magic (in the Arthur C Clarke sense of the word).
  3. Humans hate to be outperformed by a machine, but that’s what deep learning is doing in more and more domains, sparking waves of hopes and fears.

But getting ML applications out of the lab or limited application domains and into broad adoption in the real world needs a lot more. Much of the creation of this new value stack is already happening, often without the public paying a lot of attention.

So what’s the structure of the value stack of AI?

The AI Value Stack

Let’s start at the bottom:

Specialized hardware optimized for AI processing
NVIDIA and others stumbled upon this almost by accident because it turns out that Graphics Processing Units (GPUs) used for gaming and and image processing lend themselves to ML workloads almost ideally. Google and others have since produced chip designs that are specifically optimized for ML. Without a doubt, there’s a lot of room for growth here, not least for mobile devices.

Massive computing capacity made available on demand in the cloud
ML workloads are very spiky. You typically need a boatload of capacity to train models, but the actual application (inference) can be quite lightweight. That’s perfect for the kind of flexible cloud infrastructure that all the big players like Google, Amazon, Microsoft and Alibaba are already offering, but there’s plenty of space for more specialized products.

Core algorithms, libraries and frameworks

There’s plenty of innovation in developing even better core algorithms, but this segment has been disrupted by the general tendency of public and private research organization to open-source their results very aggressively. Most ML projects are now based on well-known open source frameworks such as Google’s TensorFlow or Facebook’s Torch. Of course there’s a lot of ways how these frameworks and algorithms can be improved upon, but that’s probably not a viable stand-alone business anymore. The publicly available speed of innovation is far too great.

Training data

The recently popular saying “data is the new oil” is of course very wrong in one aspect: While training data is incredible valuable and indispensable for creating sophisticated ML models, it is by no means a homogeneous commodity like oil. Small amounts of high-quality data can be much more valuable than gigantic pots of trivial, imprecisely labeled data. The best opportunities arise when high-value data is collected in smart and scalable ways, such as in the case of DeepL, a company that uses its public translation platform to assemble data that trains its best-in-class machine translation system. Just having data might not be a business in itself, but the idea of creating training data brokers that provide a thriving ML ecosystem with pre-packaged data is probably not far-fetched.

Tools and orchestration

The basic tools of ML development — such as Jupyter notebook or Anaconda — are all available for free, but there’s clearly room for more sophisticated and scalable environments. The success of startups such as DataRobot has demonstrated the need for platforms that truly boost the productivity of data science teams and ML developers.

New user interfaces

Speech-enabled user interfaces powered by ML are growing very rapidly. Amazon’s Alexa has been a huge success story, and others like Google are catching up rapidly. But obviously this won’t be the last word in how ML will enable us to interact with computers in very different ways. For example, OrCam, an Israel-based startup, makes image recognition systems for visually handicapped people and now is bringing the same technology to a broad set of consumers.

Consumer apps

Of course ML is already part of many consumer apps. The face filters that are so popular on Snapchat and Instagram wouldn’t be possible without ML. News feeds and media recommendation systems (for better or worse) have been fed by ML algorithms for years, and other everyday apps such as email and messaging are starting to benefit as well. Clearly, there are countless other ways how consumer experiences could be improved through ML.

Integrators and developers

Much like in previous generations of IT innovation, somebody has to to pull all the strings together and create complex applications for specific businesses. Without a doubt, there’s going to be a lot of space in the market for companies that specialize in these kinds of services. The big entrepreneurial opportunity will be in companies that don’t just sell billable hours, but leverage advanced frameworks to produce bespoke ML systems with much higher productivity.

Industry-specific automation

Applying ML to business problems is an almost unlimited space. Anything that requires repeatable human-intelligence tasks of a relatively unsophisticated nature is a clear target for ML automation. Even quite sophisticated tasks like the analysis of legal contracts seems to be automatable, at least partially. Clearly there are huge opportunities out there to bring tremendous efficiencies to many industries. The interesting question is how standardized these solutions can be since the environment in every company can be quite a bit different. The entrepreneurial challenge is to find the domains that have the most similarity. For example, Zeitgold is a Berlin-based startup that uses ML to parse paper-based invoices that small businesses have to pay. Automating this time-wasting, low-value process unlocks incredible value for small business owners.

Truly disruptive applications.

Thinking of Amazon as a “browser-enabled company” would obviously not even start to explain why Jeff Bezos has been so successful with his e-commerce juggernaut. Yes, the web was the key enabling technology, but countless other elements, along with a willingness to ignore conventional rules (Profits? Who needs profits?) were essential to create the currently most valuable e-commerce company.

Similarly, ML will be a transforming enabler for entrepreneurs who can really think big and outside of conventions. We probably haven’t seen these players yet, or maybe they’re hiding somewhere in a garage in rural Poland or a suburb of Mumbai.

So Where’s the Money in AI?

Clayton Christensen in his legendary book The Innovator’s Dilemma established the concept of the “conservation of attractive profits”: Whenever an industry changes structurally through disruption, profits are maintained or expanded in total, but often accumulate in a different stage of the value stack. For instance, the PC shifted profits from integrated systems (think IBM mainframes) to components such as the operating system (Microsoft) or CPU (Intel).

Similarly, the current world of business and consumer software as well as the Internet will be massively disrupted by ML technology. There’s little doubt that profits will shift to different stages, but who will be the main beneficiaries?

Right now, the Internet industry rewards context over content and specific functionality. Google and Facebook are context-providing companies above everything else (through search and social connections, respectively), while content producers often barely scrape by and SaaS companies face steep competition. As it happens, the established web giants also seem to be in an unusually strong position to benefit from the new wave of ML, not least because they have the broadest set of data.

But here’s the thing: ML does not rely on massive scale to be transformationally useful. Yes, when it comes to basic features like facial recognition, the pools of data that Google and Facebook have access to are unrivaled. But these industry giants are specialized on applications that potentially can have billions of users in the short term. That leaves plenty of space for others who are willing to start small.

Smart startups can carve out niches that might initially seem limited but can grow into global businesses. As an example: One strength of the last generation of Internet technology was easy aggregation of dispersed demand and supply. Companies like Airbnb used this capability to rapidly build platforms that brought together customers and providers that would otherwise never have met.

Similarly, creative ML startups might be able to automate industries and unlock entirely unexpected efficiencies; they could find ways to integrate computing capabilities and access to data into our daily lives in completely new ways; they might come up with revolutionary ways to drive ML performance forward; and some might (no, will) fundamentally change how things are done in the world today, making things better for everyone.

It’s an exciting time for entrepreneurs. There has rarely been a wave of innovation that has the potential to be so broad and all-encompassing as artificial intelligence. And it will be fascinating to see what all these creative entrepreneurial minds can come up with.




btov thoughts from the digital and industrial technology worlds

Recommended from Medium

System Engineering Process for AI Development

What are the most well known Tech companies doing with ML

How Much Transparency for AI Algorithms

Quantum Computing, Ethical AI, Katacoda + | Mike Loukides Interview | AI with Alex

2nd Report of November 2021

Jobs of the future: Top 5 skills that won’t be replaced by AI

Built to Last: Intelligent Cloud Technology

AI in the Healthcare World

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Andreas Goeldi

Andreas Goeldi

Technologist, entrepreneur and investor. Likes startups, gadgets, movies, good audio technology and rambling about any of those topics. Partner at

More from Medium

DeepTech startups AI strategy: data-centric or model-centric?

Age of AI — Snooping the Tech Industry, Crunching Invoices, and Reducing Clutter

AI Humility — its ok to say you don’t know

Dog jumping in snow next to a frozen Wolves shirt

Why Federated Learning can be a data sharing enabler?