Co-founder @huddle01com
San Francisco, CA
Joined July 2012
- Tweets14.7K
- Following3.6K
- Followers11.2K
- Likes78.7K
Ayush Ranjan retweeted
CONGRATS Anthropic on their S-1 filing! Unsurprising, but first observations: inflection in operating profit, strong emphasis on novel discovery/bio, mismatch b/w consumption based revenue and long term compute obligations, limited shareholder influence. a lot of case studies!
Ayush Ranjan retweeted
Neoclouds that are able to lock in forward compute prices in derivatives markets will be able to get financing as long as they have a credible track record of execution.
Pretty much.
As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers.
SpaceX closed $40B of debt for GPUs, but we’re investment grade and have Starlink & launch revenue.
That means the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many.
Ayush Ranjan retweeted
There is an absolute barrage of inference providers popping up that get to $1M/mo extremely quickly.
This is the rough economics of inference providers:
Even if you rent on-demand B300s at $8/GPU-hr, many open source models (GLM 5.3, Kimi K3, DeepSeek 4.1 Flash / Pro) can get you >$10-20/GPU-hr at reasonable throughputs (100tok/s). DeepInfer from DeepSeek directly shows they can get $38/GPU-hr on off-peak hours. Every chip / model has a different throughput / interactivity curve, but there is a lot of arbitrage between even high rental spot prices and token prices. New chips (Vera Rubins) also change this meaningfully. Your ability to be making gross margin relies on your utilization and how good your stack is.
Gateways like OpenRouter, Vercel and more help solve that problem by aggregating demand to you. May inference providers discount their tokens to get longer term commits too. I've even heard stories of selling back Chinese model tokens to Chinese labs!
One thorn in the economics of these players is that many Chinese open source model labs are now asking for rev share if you're at a certain scale, espp for pre-release access to their models for day 1 support. That further pushes the threshold to be in the money.
The barrier to entry seems low and falling. There are even stories of smaller ones who can "vibe code" a way to stand up OSS models on a cluster they rent. And as a result, we're seeing 10s of such companies pop up almost every week and immediately skyrocketing revenue at low-ish gross margin and raising rounds on the back of that.
Unsurprisingly, because the actual service provided seems commodity, this eventually drives up compute prices and drives down token prices. In theory, it would seem like neoclouds who have the lowest cost of compute and themselves have started inference businesses would win the race to the bottom (Crusoe Coreweave Nebius and GMI all have inference businesses), with larger players (Modal Fireworks Baseten) being second because they are well capitalized to make bigger compute commits cheaply and have demand (high utilization), but the reality is the pie is growing so quickly that even upstarts are absolutely crushing it and growing aggressively (inference dot net, DeepInfra, Wafer, Relace, Morph).
Strategies vary across the board: neoclouds want to serve inference, inference players want to be neoclouds and do higher margin reserved instances / post training, and some want to go heterogeneous hardware or creatively flexible compute constructs.
The risks are: do open source models continue to be competitive? Will their creators rent seek more? Will their shape change compute needs? How do different chips change economics? Who gets compute in a supply crunch?
Below is a aggressive sketch of how gross margin scales with utilization for various models / chip sets.
For now, there are many inference providers and they’re all ripping!
Ayush Ranjan retweeted
The Next Fannie Mae Is Made Of GPUs zerohedge.com/ai/next-fannie…
Ayush Ranjan retweeted
When is the ChatGPT moment for robotics?
Some say it's happened, some say it's coming. How capable are Physical AI models today? What does the adoption curve for robotics look like?
We've talked to most of the major players and make sense of it here
robostrategy.co/news/the-cha…
Ayush Ranjan retweeted
Big personal news today: I have decided to leave @a16z, move to SF, and start a company.
It's time to build :)
It has been almost 10 years since I joined the firm. I met @cdixon, @bhorowitz, and @pmarca in 2017. At the time, I was an engineer at Google Brain working on robotics. I had a deep interest in open source. I knew nothing about investing. Chris meanwhile had already planted the @a16z flag at the frontier that was early crypto with investments in companies like Coinbase. But the crypto team at the firm consisted of only him, and he had the ambition to expand. His pitch to me was simple: "You should leave Google and come help me build the crypto vertical at the firm."
The work we've done together since then to build out our team and support our founders has been the most meaningful work of my life. I am grateful to have spent a decade working with everyone here, supporting our founders, learning from Chris, Ben, and Marc, and living the firm's culture. It is a culture that values integrity, life-long working relationships, and the belief that we humans have the power to build our own future.
Marc and Ben have built a16z into an incredible force for good in the world.
Before I joined the firm, I always thought that my inevitable path was to become a founder. The truth is that I wasn't ready for it then. Somehow, this year I've gotten the overwhelming feeling that this has changed.
As for what I want to build: It is early, but I care most about building infrastructure that enables open software ecosystems. I am starting with two core observations:
1. AI today is largely a single-player technology. Most uses involve one human talking with one agent in isolation from the rest of the world. That will soon change. Most uses of AI will soon be many-to-many: many humans and many agents working together. There will be at least one massive open network on top of which those uses will be built.
2. Software is now free. It can now be written just-in-time, on the spot for a single individual's hyper-targeted use case, and it can even be thrown out after a single use. This means that our entire software stack will have to change. In the limit, it will lead to a whole new kind of AI-mediated operating systems, which in turn will create an ecosystem of AI-native developers who build applications that actively rebuild themselves on the fly.
If you are working on related ideas, and especially if you are in the Bay Area, hit me up. Let's chat!
@bhorowitz , @pmarca, @cdixon, @eddylazzarin, @guywuolletjr, @jasonrosenthal , and our whole team — I want to thank you. Working with you has been the time of a lifetime. I am forever an a16z superfan and your ally.
Onward 🫡
Ayush Ranjan retweeted
Last month, we demonstrated high-rate alchemy: turning mercury into real, stable gold, beating CERN’s record by a factor of 10.
Last week, we enriched isotopes in the world’s most powerful plasma centrifuge, beating the Soviet record that stood for 40 years.
Now, we’re building a fusion reactor in San Francisco to achieve record power density.
Ayush Ranjan retweeted
Elon Musk says blaming difficulty of bringing massive AI compute online on a "lack of electricity" is an oversimplification because power, memory, networking, cooling and construction all need to come together.
Real advantage is shifting toward companies that can actually deliver working compute rather than just announce more data centers.
Ayush Ranjan retweeted
Working with robots has always been harder than it should be.
The intelligence was hard to access. The hardware was hard to afford.
Today, both change.
@makermodsai Maker Arm + Metal Arm, powered by @dreamlabs_ai 's Dream-1. Give your robot a task in plain language, right from your browser.
Try it: playground.makermods.ai
Ayush Ranjan retweeted
It's very very different, this time.
AI Hyperscalers Net Debt/EBITDA at 0.1x vs. Internet Backbone Providers at 3.9x in 2000, a 39x difference.
The dot-com companies were funding fiber rollout with debt and capital market issuance.
Today's hyperscalers are funding AI capex internally from operating cash flow.
Ayush Ranjan retweeted
Yesterday evening I had the pleasure of attending an interesting event hosted by @sfcompute and @jumptrading focused on the compute space. Some brief takeaways:
- There's an increasing understanding that greater liquidity in compute markets requires the construction of unifying benchmark indices, so that bespoke/OTC markets can be priced/quoted as a diff to a shared set of benchmarks, similar to what exists in the energy complex already (think crude grades and power). Part of the discussion is whether this will organically emerge if more traditional players (Citadel, Jump, blue-chip banks) expand their OTC offering for bilateral hedges (using their own prices), or require an independent party (S&P, CME, etc.) to validate and give credence for broader adoption.
- Massive growth in the compute space is largely underpinned by structured financing trades with hedging mandates (similar to what you see in the oil and gas space). If someone with a sizable balance sheet just lent you a lot of money, who's likely to give you more aggressive pricing than their sales and trading arm?
- Funding from life insurers has been a key part of the data center buildout (you could call it duration matching), with an increasing number of insurers considering hedging compute. Some of this exposure sits in mortgage books, while the biggest bucket is via corporate bonds ($ORCL, $MSFT, $META, and $GOOG comprise nearly 80% of Athene's IG software debt).
- Quite a few folks seem offered over on the backend of the curve, largely driven by:
* Concerns that distressed sellers could flood the market if the AI growth story slows down (as there's a lot of GPU-backed debt to unwind)
* Natural hedgers are mostly sellers, as operators need contracted revenue to borrow, so they sell further out on the curve; buyers mostly lock in near-term rates w/ marginal demand coming from a handful of labs whose spending is contingent on their ability to still tap into financing
* Software improvements are reducing GPU-hour needs, all the while new chips are depreciating the value of old hardware per unit of compute: $NVDA is releasing a new gen roughly every year (Blackwell Ultra, Rubin, and Rubin Ultra). That said, it does look like old chips are staying useful longer than most depreciation models assume, since inference and smaller models still allow sufficient monetization of older hardware.
It'll be quite exciting to see how the space continues to mature, especially as new marketplaces begin to emerge for financially and physically settled compute. Huge thanks to @affaan and team for organizing!
Ayush Ranjan retweeted
Neoclouds such as CoreWeave and Nebius are backstopped by companies with trillion dollar market caps and access to the cheapest financing in the world
NVIDIA has a 10% equity stake in CoreWeave and has committed to buy any unused capacity through 2032, up to $6.3B
NScale, Lambda, Nebius and Iren have equity ownership and preferential vendor agreements with NVIDIA
Microsoft has signed a contract and prepaid 20% of the value at Iren. NVIDIA has backed leases at both NScale and Lambda.
Google has backed both Hut8 and TeraWulf through guarantees on the lease. Amazon has guaranteed leases for Cipher.
Yet, people still believe that these companies are at risk of default simply from looking at credit default spread charts
Ayush Ranjan retweeted
modular compute is a bet on risk
traditional DCs concentrate integration and commissioning risk at the end. modular front-loads more of that risk into manufacturing, procurement, and financing.
it's not necessarily cheaper, but it can bring revenue forward. for leveraged operators like @CoreWeave, every month between capital deployment and revenue generation matters.
and as NVIDIA standardizes much of the hardware stack, differentiation shifts toward who can finance, integrate, and deploy capacity fastest.
the next advantage in compute is cheaper risk
Ayush Ranjan retweeted
CoreWeave's debt, explained simply, because I like to speak plainly:
Secured debt = the mortgage. Backed by the house. The GPUs and the contracts they serve are the collateral.
Unsecured debt = the credit card.
Backed by a promise.
Everyone is panicking that $CRWV UNSECURED yield to MATURITY hit 15%
But a credit card rate spiking doesn't mean the house (or in this case the GPUs/data centers) lost value.
The secured lenders or the ones who underwrote the actual GPUs and contracts aren't selling here.
Now zoom out, because this is analogous to the entire AI market in wrapped up in a single company.
Trillions in secured, asset-backed debt are loaned against GPUs, data centers, power contracts, and signed capacity deals with hyperscalers. This money doesn't care about stock prices or investor sentiment.
It gets paid if the contracts pay. (Sounds pretty similar to $next if you ask me…)
Below that in line is unsecured debt, a bet on the companies themselves.
On top of everything is the equity, a bet on the story.
So when you see scary headlines about "AI debt," ask one question: which layer?
Yields spiking on unsecured debt and equities dropping while secured debt sits calm simply means the market is repricing the story.
The assets are fine.
The buildout keeps getting funded because the secured debt holders still believes in the collateral.
The day that changes… well the day secured lenders start questioning GPU values and contract quality, is the day the AI trade actually breaks and we should run for the hills.
Until then, volatility at the top of the stack (unsecured, equities) is just the shock absorbers doing their job.
