Tesla forecasts ramping up to 100 exaflops of AI training capacity by Q4 2024. That's a huge number that implies an over 20x scaling from the 4.6 exaflops disclosed at AI Day 2022 (14k Nvidia A100 GPUs), and over 50x the A100 capacity discussed in 2021. This implies a compound annual growth in training capacity of 273% from 2021 to 2024 (if they hit their target), based on the 5,760 A100s they disclosed at CVPR 21. Here are some additional thoughts / read throughs: 1. At AI Day 2022, Tesla projected to have their first Dojo exapod in production by Q1 23, scaling up to 7 exapods in Palo Alto over time. The July 23 production date on the chart, assuming they're not ignoring an already in-production pod, implies Tesla is a bit behind on exapod #1, but plans a faster, larger ramp to ~28 exapods by Q1 24 to reach capability of 100k A100s. 2. If the new capacity is all Dojo, and not Dojo + Nvidia mixed), 300k A100 equivalent performance is just under some estimates of Nvidia's total A100 SXM shipments over the last 12 months (NextPlatform estimates 350k server GPU sales of the variety Tesla is augmenting / replacing with Dojo, I'm assuming most of these were A100s as H100 started ramping recently) 3. The chart appears to be using flops as the measure of training compute. In reality, a lot more goes into end-use performance than just flops. At AI Day 2022, Tesla estimated that 4 Dojo cabinets (0.4 exaflops) could replace 4,000 A100s (1.2 exaflops) for autolabelling. This is made possible because of the increased compute density and software optimizations Tesla expected to achieve. So either they haven't been able to achieve the optimizations they thought they could, or 2024 training capability could be even higher than this chart suggests (in terms of A100 equivalent capability). Additional assumptions: - 28 exapods by Q1 24 Tesla is exclusively adding Dojo capacity after July, and hasn't added Nvidia capacity since August. The number of exapods would be lower if they plan ramp Nvidia capacity (i.e. H100s) in parallel, especially considering H100s are 4-6x more performant than A100s at AI training, per Nvidia. Sources: Nvidia Unit Sales: nextplatform.com/2023/05/01/… AI Day 2022: piped.video/watch?v=ODSJsviD… CVPR 21: piped.video/watch?v=eOL_rCK5…

Hi, I’m Frank!

From $11k per unit in 2014 to 5 cents by 2030, our research suggests data center hardware used for AI training is experiencing strong cost declines. The 48% performance-equivalent cost decline per year far outpace Moore's Law. Check out the full report: bit.ly/3gL2mUR

In proof of work, influence is dictated by possession of hashpower In proof of stake, influence is dictated by possession of native tokens Hence concentrated ownership has a significantly greater impact on censorship resistance in POS than in POW This is underappreciated

Tesla continues to build its vertically integrated AI stack w/ details on their new AI-training supercomputer, which combines 1.8 exaflops of compute with 10 petabytes of ultra fast storage bit.ly/3gPwIl6

There will likely be many access points (web, mobile, ar/vr, etc.) to what is being called "the metaverse" The real innovation is a shared data layer that offers both ownership and portability of data One company "building the metaverse" seems quite antithetical to that concept

Despite the USDC (and DAI) de-peg, the Maker protocol remained over-collateralized and fully operational through the weekend DAI in circulation is up $1 billion (+25%) Demand for more transparent, more auditable, and more decentralized financial services has never been higher.

This is wild. 600k H100 equivalents at Meta is roughly 8x Tesla's roadmap by end of 2024 (assuming 1 H100 = 4 A100). The 350k H100 target by end of 2024 suggests a more than doubling of total GPU compute since H100 started volume shipping in Q1 2023 To be fair, Tesla's relative compute increase is also insane, targeted to grow their base by ~20x in a year and a half

The growth of institutional DeFi activity as a share of total volume is shocking

Apple cracking down on NFT use inside apps is a prime example of how easy it is to forget the power central gatekeepers have today Going to be hard for NFTs to provide more functionality if the if only apps that can support them need to be sideloaded onto an Android phone

Despite overall trading volumes down m/m, decentralized exchanges just printed their best month ever relative to centralized exchanges DEX volumes ($63.7 B) are now roughly equivalent to USD-supporting exchanges ($66.5 B) Clear signs activity is moving on-chain and off-shore

I've been testing out @ether_fi's Cash card over the last 24 hours I have to admit, I think its the most impressive DeFi experience I've had in years -> crypto-native, yet user friendly, and actually real-world useful Not surprised the growth they're seeing. And awesome that its all transparent and trackable on-chain

Check out the latest episode of FYI, where @draecomino and I dig into the Terra universe with the Master of Coin himself @d0h0k1!

and so it begins…

Rather than becoming just another POS chain, maybe the most distinguishing thing Ethereum can do is remain POW (significant need for layer 2 scaling still required)

Apple is quickly broadening availability of its M1 SoC - now available in iMac and iPad Pro. Check out our take in this week's #ARKDisrupt

POW also inherently decentralizes ownership - for the most part, the largest token holders != the largest miners This is not true in most POS implementations where the largest holders also receive the lions share of issuance & fees

We published a whitepaper with Coinbase in January 2017, 8 months before they raised their series D round at a $1.6 billion valuation. We admired the company's approach to onboarding new users to the crypto economy -> leaning into innovation with high product velocity while playing by the rules. Today the company is worth ~$65 billion, representing a roughly 60% annualized return to Series D investors over the last ~8 years. Tesla, Palantir, and now Coinbase: All companies that the market fundamentally did not understand and punished as if they were going bankrupt, now recognized as category leaders in their respective spaces and part of the S&P 500. While we still think some of the greatest years for each of these companies are ahead of them, accessing innovation early, before a company makes it into the big indices, is key to building a complete portfolio.

The 200-week moving average proved resistance for Bitcoin in February. BTC traded between the 200w and its Market Cost Basis (Realized Price) for most of the month. More charts form February's edition of the Bitcoin Monthly: ark-invest.com/crypto-report…

Apple’s new M1 Max chip looks like an absolute beast

AI is going to transform how we think about software and shape the future of cloud spending. Microsoft's earnings call last night was case in point: At current growth rates, the Intelligent Cloud segment (+25% y/y) is on track to be larger than Productivity Software (+14% y/y) in the next few quarters. We just published a blog that dives into what's causing this shift and where its going 👇 ark-invest.com/articles/anal…

we thought AI training costs were falling quickly... inference costs are falling even faster! -86%+ per year fun fact: didn't even have time to add in the incremental ~43% price drop on GPT-3.5 Turbo last week More in Big Ideas 2024: ark-invest.com/big-ideas-202…

By largely leaving the scaling problem open for the market to solve in the application layer, Ethereum created competition Now the ecosystem seems to be pushing the scaling technology puck faster than any single l1 dev team could

While the new DeepSeek R-1 model provides a competitive solution to OpenAI's o1 reasoning model, ultimately, more performant open-source models are a good thing, as they provide developers with more options to build AI applications. As more AI applications scale, the aggregate need for compute will also scale. Over long time horizons, the aggregate demand for compute has only increased, and I don't expect that to change in the future. Case in point - do you think they actually faced a malicious attack this morning, or ran out of available compute when everyone started downloading the app to see why the market was selling off?

"Play to earn" is not really a new phenomenon. Here are some of my thoughts 👇

AI training costs continue to plummet. @databricks slashed their training costs by 4x in just 10 months, helped by their acquisition of @NaveenGRao's @DbrxMosaicAI. Faster than our modeled 75% per year cost decline. The pace of innovation in AI right now is unprecedented databricks.com/blog/introduc…

1/ Software has always been about boosting human productivity—automating tasks, streamlining workflows, and freeing people to focus on higher-value, creative work. AI coding agents take that to a new level.

Windows 11 looking a lot like macOS with that centered doc, rounded corners, and slide-out widgets / notifications pane

I've been quite impressed with the robustness of the Terra ecosystem, particularly penetration into payments and focus on interchain accessibility. Mirror & Anchor also gaining steam. What's the next killer dapp after that?

Something to consider when basing your thesis on a smart contract chain’s low fees: There will ALWAYS be a faster / cheaper option available. Here’s why 👇

Had a great time talking AI day and Dojo with @heydave7 and @summerlinARK!

Hot off the press! Download our latest and greatest research and tune in live to hear from our analyst team and special guests. @grousARK currently live with Digital Consumer (replays will be available)

Starting to see "login with web3" option more and more. When you own your keys, you own your digital identity.

This year’s report has the most crypto sections we’ve ever put in our Big Ideas series. Catch @yassineARK and I live in ~15 mins to go over all four of them with @bankless!

US staking-as-a-service ban rings of "China mining ban" moment for PoS networks

Resale prices for NVIDIA’s RTX 3070, the most popular GPU for Ethereum mining, continue to drop, corroborating reports of canceled orders for consumer GPUs

Proof of Work consensus secures ~80% of crypto market cap today. If Ethereum went 100% Proof of Stake tomorrow, the mix would shift to 50% POW, 45% POS.

Meta's new AI Research SuperCluster (RSC) will be one of the largest supercomputers in the world, maybe the largest With 16k GPUs and the ability to handle over 1 trillion parameter models and an exabyte of training data, this thing looks massive

Nvidia reported $155 million in sales for their dedicated crypto mining processor (CMP) in Q1. Projecting $400 million for Q2. For reference, their automotive business brought in $154 million.

To all those interested in digital wallets, payments, and public blockchains - ARK is hiring! Looking for analytically minded self starters with a passion for disruptive technology to join our Fintech / Crypto team as an associate. Apply here: ark-invest.com/job/research-…

It was a pleasure to welcome @kimbal on the latest episode of FYI. The early success of the Big Green DAO experiment is a great example of how public blockchains are inspiring and enabling novel forms of coordination - and for a good cause as well! ark-invest.com/podcast/explo…

Its funny to me that Claude does a better job of answering questions about my Gmail that Google's Gemini does. Latest versions of Gemini crush it on benchmarks and have heard annectodaly that they're great at coding, but real world use of the Gemini app has never matched up for me relative to ChatGPT, Claude, or Grok

BNB has grown from $5b to $100b+ in market cap over the past ~4 months, likely bolstered by the success of Ethereum competitor Binance Smart Chain (BSC). How has BSC gained adoption while other "ETH-killers"(remember EOS?) have failed? ⬇️ (1/7)

Competition is heating up in datacenter GPU land. Here's AMD's just-announced MI200 GPU compared to Nvidia's top of the line A100 piped.video/watch?v=ECHhuvui…

Covered DeepSeek more in this week's newsletter, particularly impressive are the details on cold-start reinforcement learning (i.e. the model learning how to reason on its own vs. from human reasoning examples)

I expect those with highest quality datasets and best tuned models to win in AI, and billions of real world miles recorded and exaflops of on-prem compute set Tesla up well for that

There are a lot of lessons to be learned from the Terra turmoil, including lessons beyond algo stablecoins & ponzinomics Worth considering: With luna worth essentially 0, there is no security behind Terra's POS network. Lido flagged this to bETH holders

Ethereum's POS validator queue is growing. With roughly 10% of the total supply of ether now staked and the queue to get in over 2 weeks, how long would it take (at least) to get to the stake rate of other major chains? Terra: 16 months Polkadot: 18 months Solana: 21 months

Had to take photos of my driver’s license for three different electric scooter apps recently, hoping Apple’s new Digital ID feature can eventually make that less painful

Does the GPU market have a hidden multi-billion-dollar inventory buildup? Our analysis shows up to a third of Nivida's Gaming revenue is potentially attributable to crypto mining this time around. Check out the details on Medium - medium.com/@fdowning/does-th…

first Coinbase, now Uniswap Labs... the Gensler era of regulation by enforcement is finally dead

Simple way to think about EIP-1559: automated share buyback program for ETH, funded by transaction fees. Why is this cool? It directly ties demand to use the Ethereum network (measures in tx fees) to value accrual in Ether, the asset, for the first time

Been forcing myself to try Bing as a daily driver search The new Bing Chat is impressive - access to open internet search makes it generally more capable than ChatGPT BUT the standard Bing search is so bad, borderline useless, that I'm switching back to Google + pined ChatGPT

This doesn't mean that Facebook, err Meta, can't succeed at building an awesome ar/vr experience that serves as an access point / tie-in to the metaverse. Oculus is already leading the pack on the hardware side, but is it in FB's dna to let users own the data?

Scaling decentralized systems is hard. Solana has been frozen in time for 2+ hours now as certain validators have run out of RAM, rendering them unable to produce new blocks.

probably nothing

Why does no one have legs in Facebook's metaverse?

Robinhood’s crypto strategy is starting to take shape. They are doubling on what they excel at, making financial products approachable and easy to use for the everyday user, and hopefully solving some hard crypto problems in the process. Excited to try this thing out.

ARK is hiring a 10x engineer to help scale the capabilitites of our reserach team. If you're an LLM and data ninja, apply here and drop me a DM with the coolest AI/data project you've worked on. app.trinethire.com/companies…

Solana just printed an all time high in weekly transaction fees. Meanwhile, Price / Fee ratio (i.e. market cap divided by annualized transaction fees) is trending down - meaning on-chain activity is growing faster than price.

Spaces should have a “playback speed” feature for quickly catching up on recordings like most podcast services AI generated summary would a nice to have that doesn’t exist widely elsewhere yet

The total stablecoin supply is now over $150 billion - growing more than 5x in 2021 Outstanding supply continued to rise throughout the year despite highly volatile crypto market conditions

With M1, the Mac Mini (a $699 device) bested AMD and Intel's top of the line chips on single threaded tasks, while holding its own in multi-thread.

There is so much in this year's Big Ideas report. Here's two graphics from the AI section that are important to key in on: Consumer AI: Personal agents are the new demand aggregators and will redefine consumer behavior Enterprise AI: Agentic workforces will increasingly digitize and scale knowledge work We think these two workflow changes will create significant opportunities by 2030 that are currently under appreciated. Check out the resulting forecasts in the full report: ark-invest.com/big-ideas-202…

I think game designers will need to think carefully about this relationship. The ones that succeed long term will likely be the ones that take advantage of the opportunities blockchains afford without forgetting that games are meant to be fun first, w/ profitable as a bonus

Yes. I’ve heard “browsers are the ultimate distribution for agents”. I think this is backwards. Agents will use browsers as a tool, among others, to accomplish tasks. You will think of the agent as the product, not the browser. You pay for ChatGPT Agent, which spins up browsers to accomplish tasks as needed. You can watch and jump in, but you don’t have to. Same direction for Perplexity’s Comet. It meets users where they are today, but over time the agent will drive more and more and you will care less and less to see how the sausage is made.

I was a bit surprised by how well Claude Code is doing. Why give up a slick web UI for the command line? But in hindsight, it clicks. Devs live in the terminal. And the CLI gives far deeper access to local context and tools than a browser ever could. It’s a great reminder: meet users where they are—not where you want them to be. No surprise that Google’s already fast-following.

Every project appending DAO to its name and launching a token is pretty reminiscent of every company appending .com to its name and going public… just saying

OpenSea delisting DAO Turtles is a great reminder of a few things: 1. The power of the UI layer: permissionless (open access) != decentralized (censorship resistant) 2. Tokens linked to a JPEG are potentially just as susceptible to securities laws as tokens not linked to a JPEG

That being said, this is a horrific destruction of wealth and a lot have been hurt, but, just like Mt. Gox & the ICO bubble, there is still massive potential ahead crypto will live on and emerge stronger for it hang in there 💪

For reference, this would be nearly 3x the size of Tesla's current AI training supercomputer Both systems use Nvidia A100 GPUs and Infiniband networking ai.facebook.com/blog/ai-rsc blogs.nvidia.com/blog/2021/0…

Had the pleasure to have @draecomino back on the podcast to cover everything going on with AI and how Cerebras is pushing the frontier of AI chips forward This was a fun one - ark-invest.com/podcast/the-c…

Anyone else feeling handicapped at work with ChatGPT at capacity? "Remind yourself that you are not alone in wanting to try out ChatGPT."

"The error in the code allowed the attacker to withdraw tokens on Binance Smart Chain when none was deposited on Ethereum." $80 million stolen Cross-chain remains one of the most exploitable parts of DeFi theblockcrypto.com/post/1321…

Assuming a median price of $250k, ETH is burning at a rate of 4 lambos per hour 🔥🔥 etherscan.io/blocks

Robinhood will need a Spotify-like driving mode soon with the number of Uber drivers I've seen checking Doge prices while driving

Mark your calendars. This will be epic!

While traditional financial institutions and centralized crypto exchanges have been barred from touching crypto in China, the country remains a top 5 user of MetaMask, the leading self-custody wallet for Ethereum

This scale reduces time-to-train and increases development velocity as they build and tune their FSD models with data collected from their fleet

Worth a read

For those who say ETH can’t scale… layer 2 is here

Not all blockchain transactions are created equal! Here are the current costs for different transaction types on Ethereum at a ~150 qwei gas price: ETH Transfer: $12.79 ERC-20 Transfer: $29.24 ERC-20 Swap: $63.97 NFT Buy: $136.90

@skorusARK the rapid improvement of fast charging for phones seems like strong foreshadowing for what’s to come for EVs

We wanted to come up with a better way to visualize all this benchmark data. So I figured I'd ask Claude 3 Sonnet. In one shot, it created a python dashboard that only took one config change to run in @Replit A few iterations (and an upgrade to Claude Pro to get around rate limits) and these are the results. IMO pretty impressive for a quick turn analysis that we look forward to expanding with more benchmarks & models over time. The outperformance in coding tasks vs. GPT-4 is both surprising & apparent (at least in my limited testing so far)

Helping train GPT-4 is probably the most productive thing Bing has ever been used for

Really exciting to see high performing mixture of experts architecture coming to open source. The leading frontier models still have an edge in absolute capability but let’s see what next year brings! Awesome work on this tracker @JozefARK.

Exaflop scale computing is on the rise, with Google's recently announced TPU v4 pod also delivering 1+ exaflops and available in GCP later this year bit.ly/3d0jy3E

75% of the Solana token supply is currently staked and frozen until the network resumes. How this number moves once things start up again will be a sign of token holder's confidence in the network. Will they keep their stake or flood to exchanges? staking.staked.us/solana-sta…

Gas prices at the pump may be reaching new highs, but gas on-chain is the lowest I've seen since mid-2020

Impressive for a first generation chip. Where will Apple put the M1 next?

Fintech + DeFI is coming for traditional banks, likely, quicker than most expect

AI could become the "assembly line" for knowledge workers, reaccelerating a decline in annual labor hours per worker not seen since the early 1900's

These violent delights have violent ends

Want to build a platform developers love and source ideas for new products at the same time? Cloudflare has a developer relations teams that’s active on Discord & Stack Exchange, helping them launch 550+ new products and capabilities last year

Nvidia's new H100 datacenter GPU is their largest generational leap in performance ever, delivering 9x more AI training performance than the A100. Cost declines at the hardware level for AI training continue to impress blogs.nvidia.com/blog/2022/0…

Easy to forget that its not just an algo stable that failed but an entire blockchain Validators halted the chain and rewrote the rules to prevent a governance takeover Maybe the right thing to do but remember this the next time you hear "decentralized"

We are rapidly approaching a world where robot token production is multiples greater than human token production Several examples on OpenAI's launch site for GPT-4o demonstrate two instances of the model conversing with each other. The customer service example is obvious: Consumer to their personal AI: call telco on my behalf to return my defective phone Telco to their enterprise AI: Answer all these inbound customer support calls and maximize for customer satisfaction and upsell wherever possible openai.com/index/hello-gpt-4…

This is a very common misconception that can't be reiterated enough. The same holds true on other POW blockchains as well. For example, minting an NFT on Ethereum uses virtually no incremental energy

Transfers of Shiba burned more ETH than OpenSea in the past 24hrs.... 🤨 ultrasound.money/