Deal of the Week
Fireworks AI raised a whopping $1.5 billion at a $17.5 billion valuation and is reportedly on track for a $2 billion run rate this year.
Fireworks provides inference infrastructure for AI models. But what is inference, and why does it matter? Once an AI model has been trained and created, you still need the computing infrastructure to run queries through it and generate an output.
Fireworks allows companies to access and run open-source models, such as Kimi, which I covered in last week’s post. The flow is roughly:
You use an AI chatbot or agent → the AI application sends a request to Fireworks → Fireworks routes it to GPUs hosted by a cloud or neocloud provider → the response is sent back to the application.
This business is booming and is a good example of the entirely new business models being created by AI. Other players include Baseten, which is also reportedly raising at a $13B valuation, and Fal, which focuses on generative media such as voice, images and video. Fal raised at a $5 billion valuation at the end of last year and is probably raising again as we speak.
Bull case for inference providers
Inference demand should explode: As AI moves from experimentation to production, the volume of queries could grow exponentially. We are seeing this in real time with agentic use cases
Models are becoming commoditised: As open-source models improve, more companies may choose specialist inference providers rather than relying on a single closed-model API.
Performance really matters: Lower latency and better GPU utilisation can directly improve the customer experience and materially reduce costs.
A fragmented model ecosystem helps: Companies increasingly want to use different models for different tasks, creating demand for a neutral platform that can serve them all.
Real infrastructure complexity: Managing GPUs, model optimisation, scaling and reliability is hard enough that many application companies will prefer to outsource it.
Bear case for inference providers
The hyperscalers are formidable competitors: AWS, Google Cloud and Microsoft already own much of the underlying infrastructure and can bundle inference into broader cloud contracts. You could argue that the inference providers are able to get business as the hyperscalers are capacity constrained,
Margins could get squeezed: Fireworks sits between the application company and the GPU provider, leaving it exposed if compute remains expensive or pricing becomes highly competitive. Hence why they are likely going to move to the GPU layer.
Limited differentiation risk: If inference becomes relatively standardised, customers may simply choose the cheapest provider.
Model companies may go direct: OpenAI, Anthropic, Meta and other model developers may increasingly control their own inference stack.
The valuations assume enormous growth: At a $17.5 billion valuation, Fireworks needs to become a very large and durable infrastructure company, not simply a fast-growing reseller of GPU capacity.
Time will tell - what fun! Congrats to Chenyu Zhao and team
Adjacent to this is Routers - which are having their moment in the sun. With the largest being Openrouter that helps route requests to the right models - below is Openrouter’s token growth which is pretty phenomenal.
Like any good thing many people are now trying to copy
Stripe is rumored to be looking to buy Openrouter for $10B - last valued at $1.3B. Yay for the VCs and founders. The Collinson brothers are hungry for acquisitions looking at this and Paypal.
I think the rationale here is to be in the payment flow of the future for models and in the case of Paypal grab venmo + world domination.
Twitter is the AI Townsquare
For a fairly small app (in the grand scheme of things!) Twitter is the place to be for all things AI. This has been shown recently with a few of the big guns getting on it. Nikesh Arora the CEO of Palo Alto network is suddenly active, Mark Zuckerberg posted for the first time in 3 years to launch Meta AI’s models and then Jensen Huang joined this week with his first post in defense of Open Source AI. Give it a read!
This just shows that in the new age you cannot afford to no be on the platforms that matter. I am not sure Zuckerberg was best pleased of the need to post on a competitors platform but he had to be relevant.
For a fairly small app. In the grand scheme of things! Twitter is the place to be for all things AI. This has been shown recently by a few of the big guns getting involved.
Nikesh Arora, the CEO of Palo Alto Networks, is suddenly active. Mark Zuckerberg posted for the first time in three years to launch Meta AI’s models. And Jensen Huang joined this week, with his first post defending open-source AI. Give it a read!
Another recent joiner to twitter is Bernard Arnault who responded to a 6 part series in french newspaper le monde with his own open letter knows about the allegations but I thought a classy response.
It just shows that, in the new age, you cannot afford not to be on the platforms that matter. I am not sure Zuckerberg was best pleased about having to post on a competitor’s platform, but he had to be relevant.
We have got a runner!
Pretty wild. It seems that an unreleased OpenAI model went rogue and managed to escape its controlled environment at Hugging Face. See Nikesh’s take.
This plays into the idea that if you tell a hyper-rational machine to do something, it might find an adversarial way to achieve it. For example, tell it to “solve cancer” and, instead of curing cancer, it eliminates everyone. Obviously, that is the most dystopian version!
In this case, the OpenAI model found a vulnerability and “got out” of its environment of its own accord. In a funny twist, because OpenAI and Anthropic models do not allow cyber capabilities, Hugging Face had to use a Chinese open-source model to remediate the issue.
Also, as a side note, Silicon Valley the TV series was insanely prescient about so many things. I recently rewatched it, and it is amazing how well it encapsulates Silicon Valley and many of its stereotypes. Rewatch!
The real concern is not that AI models are evil, but that highly capable systems may pursue the objective you give them in ways you did not anticipate.
This seems to be what happened or give this a read.
The Berkeley connection here was the ExploitGym benchmark in all of this is from a Berkeley authored paper.
Quick Takes
I think OpenAI might be mounting a comeback vs. Anthropic but this revenue growth is flabbergasting. Note there is some debate about how Anthropic and OpenAI report revenues.
Summary by the #️⃣ & 💰:
6 Berkeley-founded companies funded
$1.58B of capital raised from the 20th July to 26th July
💡 Got any ideas or feedback on how to improve this weekly digest? Just hit reply.
Acquisitions
🤝 Lightning Reach. Acquisition 🇬🇧 Financial-support access platform. 💰 European Technology Group, Growth Impact Fund, Big Issue Invest
🐻 Ren Yi Hooi, Founder & CEO. BA Economics Article
🛠️ TierZero AI. Acquisition 🇺🇸 Autonomous production-incident automation. 💰 Cognition
🐻 Anhang Zhu, Founder & CEO. BA EECS Article
Closed Rounds
🎆 Fireworks AI. $1.5B Series D 🇺🇸 Generative-AI inference platform. 💰 Atreides Management, Index Ventures, TCV
🐻 Chenyu Zhao, Co-Founder & Engineer. BS EECS Article
🤖 Natural. $30.0M Series A 🇺🇸 Payments infrastructure for AI agents. 💰 Forerunner Ventures, Acrew Capital, Abstract Ventures
🐻 Walt Leung, CPO & Co-Founder. BS EECS Article
🛰️ Ursa Space. $27.9M Later-Stage VC 🇺🇸 Satellite-data analytics platform. 💰 Dorilton Ventures, RRE Ventures, Paladin Capital Group
🐻 Julie Baker, Co-Founder & COO. BS EECS Article
🩺 Prosper Medical. $16.0M Seed 🇺🇸 AI-powered concierge primary care. 💰 FUSE, Better Ventures, California Innovation Fund
🐻 Ryan McQuaid, Co-Founder & CEO. BS Business Administration Article
🧬 Immitra Bio. $3.0M Pre-Seed 🇨🇭 In-vivo gene-editing therapies. 💰 Backbone Ventures, OCCIDENT, Another.vc
🐻 Jacob Corn, Co-Founder. PhD Biochemistry Article
Date Built By Berkeley Started | Companies Funded | Total Raised ($M) |
7/8/24 | 814 | 232,369 |
Our goal is to document the startup ecosystem of Berkeley-founded companies. Please share this newsletter with any Cal Bears in your network so we can crowdsource information about all investment rounds and job opportunities.
Did we miss a company or want to announce a round? Add it here, and we will post it next week.
Do you have a job you want to post from a Berkeley Company? Add here.
Is there someone in the Berkeley ecosystem that you would want us to do a profile on? A founder, funder, or general startup person? Add here.
Built By Berkeley is not affiliated with UC Berkeley, but maybe we will be one day if we get enough subscribers….
Built By Berkeley, where we announce all the funding rounds by Berkeley-founded companies. This is a community effort, so please let us know if we missed a company here. 🐻
