What to Watch in 2024
Some themes I’m curious about heading into next year. These aren't the VC predictions you're looking for
Who is the Azure to OpenAI’s AWS?
I’m not the first to make this analogy, but like AWS, OpenAI is ushering in a fundamentally new way to build applications. Moving data/application stacks to the cloud was/is a “digital transformation,” and companies are now embarking on a “cognitive transformation.”1
If we extend the analogy, I’m curious to see who emerges as the number two player (i.e., Azure to AWS). There are plenty of good reasons to say Anthropic, Mistral, etc., but what if the Azure to OpenAI’s AWS is… Azure?
Massive existing distribution? Check. More embedded in large enterprises and has already passed security hurdles? Check. Can bundle the product with their 365/Office/whatever it’s called Application suite? Check. Doesn’t need to engage in a race to the bottom on price to acquire customers? Check. Has a new version of Clippy? Check.
Azure’s incorporation of Meta’s Llama 2 models and release of Phi-2 reinforce this point while highlighting an emerging channel conflict between Microsoft and OpenAI. OpenAI is (currently) tied at the hip to Azure, but nothing prevents Microsoft from incorporating more models over time. Just like Azure was effectively “the enterprise cloud,” Microsoft has the potential to be known as “the enterprise AI company.” I certainly wouldn't bet against them.2
What does tomorrow's application stack look like?
Similar to the early days of AWS, the right way to build applications wasn't immediately apparent until a critical mass had made it to production and experienced consistent difficulties. Understanding those issues helped lead to products like Datadog and Hashicorp.3
So, what should we look for at the infra layer in these early days of AI-native applications? I can think of three slightly orthogonal ways of emerging as a massive player:
Build a general-purpose framework that abstracts away much of development in the cloud, but where AI infra is their wedge. This approach will resonate initially with net new applications and startups.
Develop infra uniquely capable of helping bring AI into production but compatible with existing application development frameworks. In this case, the customer base must span AI-native products and more traditional enterprises.
Infra that is low-level enough to support a future where developers have more choice on which chips they rely on. Likely lands with large enterprises to start and would expand down-market over time.
All three will likely work, but I’m most excited about the second approach because my bias in infra is generally towards companies that can demonstrate ROI (even if qualitative) within existing budgets.
BI in the AI wave
It’s easy to assume that BI is dead, and Twitter loves talking about how you can use ChatGPT to conduct analyses and create charts. But I think that misses a few things:
Uploading a simple data set into ChatGPT is not the same thing as ingesting, transforming, and centralizing all the disparate data sets required for executives to have comprehensive visibility into their business
BI is not a one-time thing; it’s incredibly iterative
Companies need standardized reporting capabilities, not repeated one-offs
Instead, I have a few hypotheses on where BI goes from here:
Data centralization matters more now, not less. I’m sure there are demos where someone can say, “Show me revenue and cash collection status by customer in EMEA,” and it will know which systems to pull data from. But how expensive (across latency, cost, and productivity) will it be to constantly pull data from disparate systems vs. centralizing in Snowflake/Databricks? Accordingly, I believe there is still a massive opportunity in data storage and management.4
The semantic layer that people have long prophesized is a unified chat interface powered by foundation models but still relies on sufficient data centralization.
The bottleneck is shifting from insight generation to communication and integration into existing workflows. The generation of insights and querying may change, but the consumption and presentation of insights won’t. If consistent, eye-pleasing reporting wasn’t critical, executives would be fine monitoring their businesses in Jupyter notebooks.5
AI-Native as a way of operating vs a product description
In Venture Land, we distinguish existing applications that layer on AI vs. “AI-native” applications built from day one using new paradigms. Startups taking the latter approach will, e.g. prioritize model interoperability earlier on and be subject to less tech debt, etc.
However, I’m more curious about how tomorrow's startups will operate. Less so the trope “teams of 10 can accomplish what teams of 30 previously could,” but instead the mindset, culture, and daily rhythms associated with it. Do you need a daily standup when an application could pull the latest PRs, check your calendar for important meetings, take assigned and active JIRA tickets, and publish that all in a concise summary to a Slack channel? Are long PRDs necessary when you can use a product that auto-generates slides and feed it a doc that has pulled a summary of your recent customer calls and mockups generated with AI? Does financial planning need to be arduous when you can use Pigment AI6 instead?
In the short term, I believe incumbents will be the biggest beneficiaries of AI. Still, I’m excited for the future when the next generation of companies adopt fundamentally new ways of operating from day one. That’s when we’ll go from wanting next-generation products to needing them.
Thank you to The Entity for this term - I’m not nearly creative enough to come up it.
What about Google? More than anyone else, Google built out (and contributed to the general public) the foundations upon which massive scale is supported, so they should have been best positioned to build cloud infrastructure, just like their contributions to AI/ML should have left them best positioned to lead the LLM revolution. But neither of those happened, and despite Bard’s impressive rate of improvement and the release of Gemini, it feels like Google has mostly lost the mindshare needed to win this market, at least at the infrastructure level.
Can’t be a VC post without plugging portfolio companies that have gone public!
If true, I’m curious if the winner will take a more generalized approach like Tabular/Onehouse or Motherduck, or be more AI-specific like Essential AI or Numbers Station.
Omni and Sigma are both tackling this problem but are approaching it from two very different angles.
Can’t be a VC post without plugging a current portfolio company!