Discover more from RealTech Weekly
🔥RTN20: From Big Data to Big Lender
SunRun earnings, DeepMind's new robotic model, new Tesla acquisition
Evolution of capex financing?
Coreweave, this week announced a $2.3bn debt facility, collateralised against its Nvidia chips. Coreweave is a cloud HPC provider, that rents access, predominantly, to Nvidia GPUs (you can see their pricing here). As you know, AI companies need an obscene amount of GPUs to train SOTA models;
This has created a ‘GPU arms race’. Nvidia chips are selling for 1.5x times their MSRP. This inflation is driven by huge AI demand and supply bottlenecks at TSMC. Large funding announcements make a point to highlight GPU acquisitions:
Coreweave’s debt financing (secured against recoverable assets) is the evident form of financing scaled, CAPEX-intensive businesses. Whilst I don’t know the details of the deal, I would speculate at a 25% to 50% Loan-to-Value + equity, secured against cashflows, then Nvidia GPUs.
Buying ‘commodity’ hardware is not a great use of risk equity capital for a scaled business.
Yet this deal has been viewed as novel in the tech world. Why?
There has been a relative lack of sophistication within early-stage venture funding. Primarily driven by low-interest rates, cheap and abundant equity across all stages and a lack of understanding of cap structures by VCs. Though, this is going to change.
Venture Capital will be concentrated on investing earlier on the risk curve (pre-Product Market Fit), companies will drive to profitability earlier and maintain higher funding optionality
As we enter a hardware and capex-intensive cycle, there will be an increase in non-equity capital available to startups to fund growth such as;
Asset-backed lending: borrowers will lend against hard assets that have recoverability value, where they can understand and underwrite depreciation schedules. These might be chips, plant facilities, AR, net operating losses, inventory etc.
Sale and leaseback: Prevalent in RaaS (robotics) - vendors sell their hardware to a lender to generate cash upfront, the company leases the hardware, taking a spread
Project financing: The company contracts with a customer or government to deliver the project under a concession or offtake agreement that commits future revenue
As we covered last week, there has been a claim around a new superconductor
There has been a race to replicate the experiment in order to validate its claims
This week, saw a number of claimed replications of the experiment, with some being successful at -167 Celsius at ambient pressure.
Chinese researchers with the Huazhong University of Science and Technology (HUST) have claimed to have successfully replicated the superconductor's manufacturing process, posting a video on Twitter
Andrew McCalip at Varda Space, posted this video
Lawrence Berkeley National Laboratory (LBNL) replicated the synthesis in simulation, as did multiple other labs
Physicist Andrew Cote’s Twitter is likely the best source to stay tuned to developements
What is clear, is that there is a palpable excitement from the scientific and startup community about the implications of LK99
Which, could, affect every industry that uses electricity
SunRun is a leader in US home solar. They sell solar panel systems to homeowners, typically through a long-term lease or power purchase agreement
Total revenue was $584 million, up 30% year-over-year.
Customer agreements and incentives revenue was $233 million, up 20% year-over-year.
Solar energy capacity installed was 296.6 megawatts, up 20% year-over-year.
Net subscriber value was $12,321 in Q2, up from $12,000 in Q1. This metric excludes the impact of IRA tax credit adders which will benefit margins in the future
Leading in grid services, with the expansion of virtual power plant (VPP) program with PG&E to 8,500 customers
Why it matters: RT-2 combines the benefits of web-scale pre-training, efficient fine-tuning, and a powerful VLA backbone to achieve stronger generalization, emergent skills, and sample efficiency compared to prior work. This shows that LLMs can be used for training where there is sparse data available, such as robotics
RT-2 is a novel vision-language-action (VLA) model that learns from both web and robotics data, and translates this knowledge into vastly improved generalisation for robotic control
Leverage a pre-trained Vision-Language Model (VLM) like PaLM-E as a backbone model. The output tokens represent discrete robotic actions instead of text
RT-2 shows impressive emergent skills like symbolic reasoning and human recognition not seen in RT-1 or RoboCat
Best of the Rest
German chip maker Infineon is building a Fab in Malaysia (!? not sure why they don’t build it in Germany, EU sovereignty?)