AI’s Broken Staircase

How startups can accelerate industry adoption of AI

Isabelle Zhou
5 min readDec 19, 2023
Image: DALL·E

It’s no secret that ChatGPT, Claude, and Poe have exploded into public awareness this year. Smart, reactive, and occasionally snarky, they’ve gone brazenly viral across the globe.

In the last decade, artificial intelligence’s potential has drastically expanded through a nexus of scientific breakthroughs, data infrastructure, and increased computing power. We’ve seen a few inflections:

  1. Computer vision brought us facial recognition technology, augmented & virtual reality, and self-driving cars.
  2. Machine learning classifiers gave us language translation, social media algorithms, and better medical diagnoses.
  3. Generative AI redefined our day-to-day reality by letting us co-create with AI tools, embrace smarter workflows, and even form real relationships with bots.

As quickly as AI technology has evolved, we are even more quickly approaching the limits of AI adoption across industries. Today, there are 4 new broken steps that need to be fixed before the world can ascend into a future that fully harnesses the potential of AI innovation.

🪜 Step 1: Compute

At its bare bones, AI = compute. Artificial intelligence works by running large sets of data through sophisticated computer algorithms to teach computers how to classify and understand this data. The bigger the data set and the more sophisticated the algorithms, the better the output. We see that computing costs scale exponentially with better AI:

Chart: The Economist

Previously, AI models like OpenAI’s GPT-4 got smarter by training on more and better data. Now, we are rapidly approaching a broken step where we physically cannot build data centers fast enough to accommodate this volume and where each new piece of data has diminishing marginal value.

As the AI race heats up, computing will be a critical resource constraint. While incumbent players such as AWS and Azure in cloud computing as well as NVIDIA and Intel in silicon chips currently supply a lion’s share of the world’s computing power, there is room for startup innovation in how these computing resources are distributed.

Startup Opportunities:

  • Serverless computing businesses that enable AI companies to automatically adapt to computing demand & usage, and to minimize service outages which have become common.
  • Businesses that sell flex usage models to startups, which decreases the barrier to entry for AI startups that cannot afford dedicated server space from traditional providers.
  • Sustainable energy providers that optimize electricity costs as the energy needed to power AI models becomes exponentially massive.

🪜 Step 2: Data Tools

AI’s “intelligence” is only as good as the data we use to train it. If we want to step up AI’s performance, especially for specialized tasks across diverse industries, it becomes mission critical to not only gather data at scale, but also efficiently label and clean it in a way that will make it palatable for training machine algorithms.

Startup Opportunities:

  • Data labeling businesses that innovate on technology or operations. Much of data labeling today is done manually, often outsourced to cheaply paid data labellers.
  • Plug-and-play tools that scale algorithm training such as through enabling weak-to-strong generalization and AI superalignment
  • Vector search and data filtering businesses that can quickly grab the most relevant data subsets and moderate out unsavory data.
  • Data collection & organization businesses that can creatively gather relevant data across the web and physical worlds. Most of the world’s data today is unstructured and lives outside of formal databases.
  • Businesses that enable other digital and physical businesses to monetize their latent data as an additional revenue strategy.
  • Privacy-protecting businesses that allow AI products to use but not store personal data; this enables highly customized user applications.
  • Data marketplaces that allow companies to filter and purchase certain subsets of data. If proprietary data ends up being a store of value for AI defensibility, then we may see a whole new data economy emerge.

🪜 Step 3: Developer Infrastructure

To truly scale industry-wide adoption of AI, we need better and cheaper tools to enable any business to put the latest developments in AI to use in real life. While AI infrastructure tools are improving every day, there are still strong technological and financial barriers to entry around using and monitoring AI models. With better developer tooling that reduces the friction to spin up a generative AI app, more businesses can be encouraged to build up AI-enabled workflows.

Startup Opportunities:

  • Businesses that offer automated model optimization and deployment to reduce the time to run multiple trials of a model.
  • Businesses that accelerate collaboration between data scientists, machine learning engineers, software developers, and IT ops teams.
  • Businesses that offer the ability to quickly and automatically run data through machine learning models to provide simulated results and fine-tune custom weights and biases.
  • Horizontal platforms that dynamically connect frontend queries to the right model endpoint (assuming the product runs off multiple models).
  • Standardized developer interfaces that can facilitate LLM calls and run parallel model instances to reduce downtime.
  • Security platforms that can secure data and API endpoints, keeping them siloed to each user or potentially erasing them after use if sensitive data is handled.

🪜 Step 4: Foundation Models

While large companies such as OpenAI, Anthropic, Meta, and Google have launched popular general foundation models, there is still opportunity for smaller startups to launch more nimble models that can stand out by inserting cleanly into specialized workflows, owning vertical-specific use cases, or pushing the frontiers of multimodal output.

Startup Opportunities:

  • Sector-specific businesses that train a large volume of proprietary data for niche use cases not easily serviceable by large generalized models.
  • Businesses that innovate in their foundation model output format (datasets, structured texts, programmable lists, bot actions) and plug neatly into industry workflows.
  • Businesses building models that ingest and respond to multimodal inputs, such as those driven by movement, haptics, eye tracking, etc.
  • Businesses that focus on non-English prompts, with stronger adaption for cultural contexts than current state-of-the-art models.

There is a real opportunity right now for startup-led innovation to help build the next steps up the staircase of broader AI adoption. We are only at the beginning of this climb and a new AI-powered world will have ripple effects touching every aspect of our lives.

This piece was written during my time at Floodgate. Special thanks to Arjun, Mike, and Ann for their support on this piece. The Floodgate team is incredible and can be reached at if you’re an early stage founder.

Please feel free to reach me on Twitter @isabelle_zhou



Isabelle Zhou

Formerly investor @Floodgate, founding team @Nooks AI, cs @Stanford.