Why the AI Investment Opportunity is Different
On the one hand, the leap forward that AI took last year is real and opens the biggest opportunity in technology since the development of the web or social media. On the other, unlike previous waves, everyone is aware of it, not just the experts, techies, or early adopters.
The net result of this is that although the signal indicating the opportunity is loud, the overall signal-to-noise ratio is low. This will result in an investment landscape that will be tricky unless it is put in context of the structural change it creates and capital is deployed with a deep understanding of the details.
Putting AI in Historical Context: AI as Third Phase of Internet
AI has been around for decades, arguably since the development of computing, but the release of ChatGPT in late 2022 represented something transformational — a phase shift, caused by progress in generative AI foundation models, reminiscent of how the growth and utility of the internet exploded after the development of the web.
If the development of the web represented the first phase of the internet in terms of investment driven commercial activity, the second phase, dubbed web 2.0, evolved from the fusion of RSS aggregation and blogging to create a single feed of aggregated content from followed accounts combined with an effortless way to publish online content. This many-to-many, read/publish model is the one that underlies all social media and, if the dotcom era web was ‘read-only’, web 2.0 was ‘read-write’. It is no accident that Twitter was created by the founder of Blogger, and I was able to witness this firsthand, having co-created a hybrid web feed blog posting service for Blogger and co-authoring the RSS standard.
There are many candidates for further phases of the web, such as the development of mobile or of internet meets physical services such as WeWork, Uber and Airbnb, but in terms of a third pillar, a ‘web 3’ to complement web 1 (the web) and web 2.0, it isn’t the metaverse or blockchain-driven decentralization or what has recently been labelled ‘web 3'. The third pillar is AI.
In many ways the moniker web 3 is a label that refers to a failure rather than the third wave of the internet. It was formed around the fusion of the worst of the crypto/blockchain/DLT/De-Fi (pick your label) space, sometimes tacked onto the notion of the metaverse and Facebook’s quixotic bet on VR (which Apple has alarmingly doubled down on with sheer hardware, brute force).
Crypto and the broader fintech space crashed (both drops were worse than the dotcom one which followed the first phase of the internet). The blockchain world was still struggling to find use cases for decentralization beyond the brilliantly pure one of Bitcoin, 14 years after its inception. Interestingly the previous candidate for the rebound of the internet after the dotcom crash was also around decentralization, peer-to-peer, and file sharing. The biggest of these decentralized services, Napster, was created by the person who later became president of the most centralized application of all — Facebook, as social media not peer-to-peer decentralization became the focus of activity for web 2. In fact, the label web 2.0 was coined by the organizer of the peer-to-peer conference which rebranded as web 2.0.
Yet again, the next wave of the internet is clearly AI not the decentralized ‘web 3’.
AI In Computing History Context: AI as Alternative to ‘Code’
If AI represents a third internet wave, what is it that is structurally different?
The web and subsequent dotcom era were founded on the emergence of http, websites, and the web browser. It was largely a read only medium, where creating a website was much harder than posting to social media, but reading was much easier than before, since hypertext links allowed for an endless random browse through multiple websites. Social media transformed the web from read to read/write and this was the structural shift that warranted it being considered a second wave of the internet.
The revelation that AI is a new step change in general use computing capability means that there is a similar level of structural differentiation as there was between the first and second phases of the internet. This time the change is in the way that software itself works rather than how it is interconnected. Until now, nearly all the world’s software operated by encoding processes into Boolean logic “if this, then that”, etc. It meant that ideas and processes needed to be codified or understood in terms of pure logic, for them to be translated into software and as a result computing was the realm of logical, robotic, deterministic processes. Conversely, AI is based on a network of non-linear functions which create nondeterministic outputs from weighted inputs. Modern, machine learning based AI is predicated on settings that work to produce valuable outputs, but we do not know how, and the results are probabilistic and human-like in both their flaws and benefits.
This fundamental shift between a world of codified logic to one of black-box, emergent capability means that the core capabilities of computing have fundamentally changed, and so vast new areas of opportunity and investable use cases will open up.
One way to look at this is that software itself, based upon neural networks, has become a network and that this creates similar structural changes to business models that the web unlocked.
The web allowed internet applications to be created which had completely new and more scalable distribution allowing for business models based on capturing virality and network effects — these being the principal characteristic benefits of connectedness. This created an opportunity for some internet companies to grow faster than their industrial predecessors, such that they could capture an entire market segment and wipe out most of the competition. This happened with Google and search, Facebook and social media and Amazon with retail, to name just a few.
New internet driven business models meant new investment models. The risky bets needed to back the winners in winner-takes-all, internet platform opportunities, established the Venture Capital model of investment as the default one that underpinned the shift from the industrial to digital era. Venture Capital allowed the increased risk (only one winner from lots of contenders) to be offset with uncapped upside from equity in a way that traditional debt-based investment did not allow.
We do not yet know what the exact investment dynamics are for capturing AI opportunities, but we know that it will be quite different from the current status quo and the drivers of this difference will be in its position relative to the structural differences between each previous phase of the internet, I.e.
Phases
Phase 1, the web. Network enabled information browsing via hypertext.
Phase 2, web 2.0. Network enabled browsing AND publishing via social media feeds.
Phase 3, AI. Software internals as a network and potential self-replicating software development.
Structural Differences
Phase 1 & 2 structure: networks which connect software applications and people using them.
Phase 3 structure: software whose internal workings are a network.
Benefits
Phase 1 and 2, distribution network benefits (virality and network effects) allowing for low customer acquisition costs, exponential user growth and moats and monopolies through network effects.
Phase 3, production network benefits allowing for exponential software improvements and capability leading to efficiency gains in productivity and exponential increase in functionality.
How it plays out
One possibility is that AI will move digital era network benefits from distribution to production for lock-in and runaway productivity gains that are impossible to compete with, creating insurmountable moats and sustained monopolies. Examples of how this could happen would be if, say, OpenAI were to win a race to AGI (general intelligence) and have GPT create better versions of GPT which in turn create better versions of GPT etc. Another example would be if vast sales of Teslas created enough video for true self drive, allowing Tesla to sell cars at cost, wiping out competitor OEMs and charging a premium only for their software. Given that Tesla profits are coincidentally tracking software and services revenue, this is undoubtedly their game plan.
Implication
Internet investment to date has not really been about technology but sales and marketing advantage through distribution network benefits. Going forward, AI will favor investment in potential capabilities and production network benefits, and this will look a lot more like investment in software itself. But because AI systems are a black box, the capabilities will just need to be demonstrated and not necessarily understood how they have been achieved. In this sense ‘tech’ investment will still not really be about investing in technology and still about network benefits, but production rather than distribution network benefits.
Areas of Investment Focus in Layers of the AI Ecosystem
Having established the historical context and structural difference of the AI era internet, what does it look like in terms of components — what is the AI internet stack, who benefits and where should investments be made?
At the lowest level, chip manufacturers such as Nvidia have been the biggest beneficiaries, providing the hardware for the massive amounts of compute power that AI model training and prompt processing require. Likewise, the incumbent tech platforms (Google, Amazon and Microsoft) are beneficiaries of the AI revolution, via their cloud services offerings but the outright winner here depends on whether LLMs commoditize or not.
We are pessimistic about Apple as its AI offerings are weak compared to others and it has focused on VR/AR for the single new piece of hardware since Jobs’ death that is not a peripheral device to the iPhone. Apple’s stock price performance is largely based on efficiently monetizing and augmenting existing innovation and returning cash via stock buybacks The strategic direction of the Vision Pro headset feels like wishful thinking in terms of the next ubiquitous computing form factor to go beyond the smartphone.
Google or Microsoft look closest to benefit from a real challenge to OpenAI’s head start in foundation models. Meta has a potentially unique advantage if its open source LLMs commoditize the entire foundation model layer around an architecture and ecosystem that it controls and move revenue opportunities to above the foundation model layer itself (tools for creating fine tuned models) and below (cloud hosting).
In the enterprise application layer there is a unique opportunity for AI to be different from previous waves since the capabilities and potential business applicability shown by products such as ChatGPT are so obvious. And because these benefits are communicable to C-level executives within large corporations, corporate check books are open far earlier than they were in the dotcom or social media era. With enterprise sales cycles shortened and clear use cases such as LLMs which are fine-tuned on proprietary, siloed, corporate data, startup revenue comes more rapidly, and markets remain fragmented based on who can do a sales land grab. This creates a less winner takes all dynamic at this layer and with earlier profitability means less VC cash required to reach profitability. This environment may exhibit regional clustering, corporate venture investment and bolster some incumbents.
At the consumer application level, we don’t yet know and capital (at least in the US) has been allocated too quickly and with not enough discretion. What we do know, however, is that AI interfaces will be very different and will be what we at Aperture are calling ‘semantic interfaces’. In semantic interfaces, AI software can be instructed by delivering the meaning of what you want to achieve rather than stepping through a bunch of commands that relate to sub steps with no self-contained meaning. To explain this, consider an image processing application such as Adobe’s Photoshop. In the past, removing clouds from a picture with some sky would involve many steps using pen tools and masks and filters. With AI we can create menu items that simply say things with specific meaning such as ‘remove clouds’.
Summary
AI is the real third wave of the internet. What was called web 3 was a failure and is different. The AI era represents a shift from code to settings and from deterministic to probabilistic computing. If the previous waves were about networked applications and users, this is about applications themselves being networks, ‘infranets’. At a business model level, the structural changes will play out in the production rather than distribution advantage, and this means investing in product capability rather than viral sales and marketing potential. From an investment perspective it creates different opportunities at different layers in the AI ‘stack’ but for very early-stage investors the current sweet spot is around simple enterprise applications.
