Who will win in the next phase of AI?
“Any sufficiently advanced technology is indistinguishable from magic”
Arthur C. Clarke
There is now very little doubt that AI will be the next major platform shift. But who will be the winners in AI? I’ve sketched out some hypothetical scenarios and the conditions required for each type of player to win. While this is highly speculative and it’s often a futile endeavour to make predictions, it’s still a fun thought experiment.
Here’s a tour of some potential players and the scenarios for them to win. For good measure - not investment advice, DYOR.
1) Incumbents/large corps (Microsoft, google)
Scenario 1: Building these AI models continues to require proprietary tech and prohibitively high cost of compute. Capability compounds and leads to exponential performance growth (much like silicon chip production). Large tech moat forms around models, smaller startups cannot compete on performance and die out. Winner take all/most scenario.
Scenario 2: Machine learning (ML) model architectures become ubiquitous and commoditised. Access to vast amounts of unique training data (esp real time) in a closed ecosystem becomes a differentiator to output of models (transaction, social data). Existing networks dominate (meta, amazon etc).
In either scenario, proprietary tech leads to an unstoppable takeoff. Other players are at the mercy of these AI giants and they can command high prices for API usage. Duopoly or oligopoly scenarios (few large corps develop models) could introduce competition and drive costs down. Overall massive value creation to ecosystem. Bad long term outcome: large incumbents use superior AI capabilities (using generative AI to create functioning businesses) to displace players upstream and downstream, leading to the first true multi-industry uber-monopolies.
2) New startups/ventures with proprietary ML models
Scenario 1: A new startup creates a breakthrough in deep learning architecture (beyond transformers) and exceeds all existing competitors either in generative AI space or unimagined domain. New models are not open source. Similar outcome to above.
Scenario 2: Transformer architecture remains the primary breakthrough and new deep learning techniques are mostly open source/transparent. Instead of building new large language models (LLMs) from scratch (very expensive and computationally extensive), companies feed niche data to existing models, topping up the unique 1% data on top of the 99% existing model (Sam Altman’s analogy). Ventures that develop a uniquely valuable model with a strong use case within a niche dominate, leading to multipolar scenario with specialised platforms for different industries (ex. protein specific data for biotech models like Alphafold). Novel models get created and applied to use cases we’ve never considered. New industries emerge. Market grows larger but remains relatively fragmented.
3) New “layer 2” companies with APIs
AI companies with models become commoditised and accessible through APIs, triggering a wave of startups that plug-in to AI models to provide new services and products. (OpenAI already lists 300 companies using its service). ML models become backend infrastructure, powering the wave of innovation but with large gains going to downstream companies (think AWS providing cloud services to Netflix) that successfully leverage and commercialise it (solve specific problems, better UI, develop network effects).
4) Existing players that leverage AI
Think existing, major (non-AI) companies, from lifestyle to finance that plug-in to LLMs (already happening, Notion + openAI, copilot). Similar to the new layer 2 companies, they utilise the wave of new advances through integrations, creating immense value with their existing customer base (Stripe, duolingo integration with GPT-4). Companies can plug-in a “universal autocomplete” equivalent, shaving down development and process time to a fraction and reducing costs by over 10x. Generative AI goes on overdrive and existing large companies spawn new verticals within their enterprise to become behemoths.
5) Pick and shovel
When everyone is competing to extract gold from the same gold mine, intense competition could eat away at surplus and leave only marginal profits. Sometimes the players that supply the tools end up the big winners. In this case, the shovels can be the GPU (chips optimised for machine learning calculations) producers and cloud computing companies that will massively benefit. Higher demand will spur more investment into hardware which will increase computing power (more innovation, more demand) in a positive feedback loop.
6) Singularity/Intelligence detonation
Someone invents a learning model that can recursively improve itself. Exceeds human level intelligence within an hour, superhuman with the next half an hour, and god-like inconceivability within the next few minutes. Depending on initial goals set, we might all end up as paper clips.
Future
The outcomes (excl singularity) generally fall into two camps: winner take all (monopoly/small oligopoly) or multipolar/fragmented. If history is any indication, the power law will prevail and we will be left with a small segment of ultra profitable winners, followed by a long tail of companies that go extinct or survive on the margins. Unless a single monopoly takes over, all the other scenarios are not mutually exclusive.
My short term prediction is that we will see an explosion of “layer 2” startups that leverage existing ML models (can move faster than existing corps, easy to get funding), more innovative corporations that immediately plugin (stripe etc), and AI incumbents like Google unleash a onslaught of AI tools in their ecosystem (Google workplace) There will be over-investment due to fomo (similar to 1990s) followed by pruning over the years as dominant players emerge. Or outcome 6 could happen. Who knows. The future is exciting.