StabilityAI's Recent Fundraising In Context
Sandia's Thunderbird Linux cluster Jet Photo: Randy Montoya / Sandia Labs
Stability AI raised $101 million this month to fund greater computing capacity and various other operational costs. This funding will back the company's efforts to provide open-source generative AI models, such as the firm's recently successful Stable Diffusion text-to-image model.
While such large fundraising rounds were uncommon in previous generations of tech startups, they have increasingly become the norm for recent AI-oriented businesses. For example, Peter Thiel’s famous investment in Facebook was only $500,000. These relatively small investments were successful because of the low capital costs associated with Web 2.0 startups. However, the current generation of AI businesses requires far more investment. Stability AI’s CEO has claimed that next year Stability will catch up to Meta’s computing ability, at least in terms of raw numbers of A100s and H100s (NVIDIA’s latest AI-oriented chips). This claim makes the need for large investments clear, with Meta’s market capitalization still exceeding $350 billion after the recent downturn.
However, incumbents in the tech world will not be surpassed without resistance. Earlier this year, Meta built the AI Research SuperCluster, bringing them to the forefront of the private sector’s effort to build huge computing capacity. Google’ computing capacity is harder to estimate because they rely on custom Tensor Processing Units (TPUs), which have different architectures than their competitors' A100 and H100 chips. Still, their investments remain in billions each year. Finally, OpenAI has built one of the largest supercomputers of the past few years in partnership with Microsoft. This trend towards private companies building their own native supercomputers is unlikely to let up soon. NVIDIA’s H100 chips are set to start reaching data centers soon, with some projections estimating that they will drive NVIDIA’s data-center sales to surpass their revenue from gaming.
Until recently, the world’s biggest supercomputers resided in government-funded labs. For example, Oak Ridge National Laboratory debuted this past summer and took the spot for the fastest supercomputer in the world. Similarly, Japan poured $1 Billion into the Fugaku supercomputer in Kobe, Japan. Researchers have used these systems widely to conduct basic scientific research on various topics, such as the effects of prescription drugs and the projected fallout from nuclear detonations. Most of the uses of the world’s largest supercomputers revolved around providing simulations to highly complicated systems, a task which was not extremely valuable to the private sector. This was true until breakthroughs in machine learning opened new profitable uses for supercomputers. In the past couple of years, there have been huge breakthroughs in AI technology for natural language processing, text-to-image generation, and even Minecraft gameplay. These breakthroughs follow new approaches to AI, which emphasize the positive relationship between the size of a model and its performance. With a plethora of large companies focusing on building better AI models, the capacity to train huge models is often an essential differentiating factor.
A NVIDIA V100 built for Tesla / NVIDIA
These shifts in market dynamics also bring into question the balance of global power between countries and private entities. Recent restrictions on the flow of critical semiconductor technology have brought to attention the various misuses of AI technologies, such as guided weapons and surveillance technology. They have also demonstrated the will of the US government to implement restraints on the private sector. Currently, there are no significant efforts to slow the arms race towards technological supremacy in the private sector, but similar restrictions may be placed on companies that amass too much power in the future. OpenAI and DeepMind have made it clear that keeping increasingly general AI systems in line with humanity's intentions is at the forefront of their missions. Perhaps computers that make powerful technologies increasingly easy to access will eventually warrant legal regulation.