It's pretty crazy that right now, the highest margin thing to do with these models seems to be simply selling them through an API. Dwarkesh's blog prize[1] questioned how this dynamic could ever result in lab profitability, simply because the scale of reinvestment into model training and research requires constantly reinvesting more than you're making. Well, Anthropic is likely already profitable,[2] and it hasn't required any of the schemes I saw proposed in answers to his question. It turns out that the margin on selling frontier intelligence through an API is just really high! So high, that AI training and inference is consuming supply chains that used to serve other high-margin industries. Memory and fab space is being redirected from consumer devices,[3] and compute that used to serve Bitcoin mining is being repurposed to capture a slice of those frontier lab margins.[4]
This dynamic reinforces classic concentration of power risks, with some estimating OpenAI and Anthropic will soon control over 80% of worldwide compute.[5] As long as supplying frontier labs is the highest margin usage of compute, it's hard to not see this becoming the case. However, I think there is a decent case to be made that trading will become a threat to research compute, in the same way that research is now a threat to compute in consumer products, due to better profitability and the rapid growth.
Trading firm AI use
Hedge funds and trading firms have used "AI" for decades, as that amorphous label can apply to everything from algorithmic market making to high frequency trading. I want to focus on a specific trading strategy that seems to be gaining traction in that industry: compute intensive mid-frequency trading. As a reductive summary of the many different ways this is employed, this essentially involves training a model to predict the direction of a security over a short timescale, maybe 5 minutes, and trading on that assumption. The details of the specific trading strategies and data sources are closely guarded, and often target different asset classes or markets firm to firm (options, public equities, bonds, etc.). The compute needed to perform trading strategies has long been pretty meager. At the start of 2024, the most compute intensive player (Jane Street) had committed to only around $250 million in GPUs, and the MFT space as a whole was likely less than a billion.[6] These numbers don't touch the compute commitments of the labs and wouldn't be worth including in models of their economic power. That has changed.
A note on the numbers. Due to the secretive nature of trading, many of the numbers I'm throwing around are best estimates backed out of public announcements. These firms have incentives to both downplay and emphasize the success of certain strategies, which limits the ability to perform analysis like this. The people with the most visibility here are the firms themselves, so this is mostly based on their committed actions and disclosed returns. I also don't personally have a trading background! I come at this from an AI policy perspective, as I'm worried this risk might be overlooked.
In January of 2024, Hudson River Trading and Jane Street had around $600 million of contracted compute capacity combined. HRT's was mostly on-premises through an installed base of H100s, and Jane Street had just begun an agreement with CoreWeave to access a few thousand of their GPUs. This level of compute, even if paltry compared to the frontier labs, represented the lion's share of GPU capacity across Wall Street. The only exception might be XTX Markets, who had a significant amount of GPUs (25,000+!). However, they were focused on using it for specific low-latency market making strategies.[7] How HRT and Jane Street were using their compute was more unclear. In 2023, most of HRT's revenue had come from high frequency trading strategies and likely pulled in around $4 billion in revenue that year. Jane Street generated around $10 billion from a mix of approaches. In 2024, they both doubled revenue to nearly $8 billion and over $20 billion.[8]
These firms might simply have been cashing in on high levels of volatility. We can roughly compare their returns to other similar firms without compute. Citadel Securities' net revenue was about $6.3B in 2023 and $9.7B in 2024, increasing 55%. Virtu Financial had revenues of $2.29B in 2023 and $2.88B in 2024, up 24%. Optiver grew around 30%.[9] These firms are growing sizably and seem to be taking advantage of that sustained volatility, but none of them approached the growth seen by HRT or Jane Street that year.
Jane Street and HRT from bondholder disclosures reported by Bloomberg; Citadel Securities per Bloomberg; Virtu adjusted net trading income from SEC filings; Optiver from its annual results. 2026 is first-half revenue annualized (which honestly might undersell the trend) and for Jane Street Q2 is estimated. Lots of assumptions and napkin math, but the general shape should be correct.
Whatever caused this growth, both Jane Street and HRT started taking compute much more seriously after 2024. Jane Street announced expansions to tens of thousands of GPUs in the first half of 2025 and secondary sources reported that HRT was on track to spend a billion on compute that year.[10] Between the two of them, they ended up spending over $2 billion. No other firm in the space had disclosed any material compute spend. Reports came out that the companies were trying new kinds of trading. HRT in particular had mainly been known as a high-frequency trader and suddenly they were holding assets on their books longer to perform this mid-frequency trading.[11] On an Odd Lots podcast in late 2025, they directly attributed massive compute usage for the success of these kinds of strategies.[12] Again, both firms grew their trading revenues more than 50% on the year. Still, no other sizable competitor for trading compute had emerged.
In the first quarter of 2026, both firms pulled in more revenue than they had all of 2023.[13] Then the second quarter came in around double the first.[14] Both these firms began scrambling for compute; Jane Street signed a $6 billion commitment with CoreWeave in April, while HRT bought a billion of B200 capacity.[15] After Jane Street's $16B first quarter represented 40% of the previous year's revenue, heads started to turn on Wall Street. Previous anecdotes from firms like Citadel had dismissed many ML strategies to hold equities as "falling apart" when predicting further than 5 minutes into the future.[16] It sounded like Jane Street might be using models to recommend holding on the scale of days or even weeks.[17] Other firms followed their lead. Flow Traders, DRW, and IMC committed to purchasing a billion in compute capacity, from owning essentially none previously.[18] Jane Street and HRT had now both grown their revenues around 6x in the last two years, compared to less than 3x for any peers without meaningful compute.
By September, HRT and Jane Street already found themselves, again, compute constrained. Jane Street signed a deal valued at $13 billion over 5 years with Crusoe and HRT paid CoreWeave another few billion for capacity.[19] Of the around half a gigawatt of compute committed to trading, these two firms still control 80% of it. On the .15 GW in use this year, the firms could potentially generate over $150 billion in revenue. Again, exactly how much of this revenue is directly attributable to these strategies and how scalable they are with more compute is unclear, but the people who would know best are buying a hell of a lot more.
some of these deals are vague/multi-year but only show up in the quarter they were announced at the total contract value, which makes attribution and good modeling hard
One relevant piece of academic literature[20] suggests that this is not sustainable; that an increase in AI trading strategies will erode the alpha of all using them because of the correlation in the resulting trading behavior. This could very well happen, but a key difference with general AI trading strategies is that the models are not exploiting a specific market inefficiency. Rather, they would be superhuman at finding new inefficiencies to exploit. The models could, as well as the people writing that paper, understand how their signals would correlate with other AI trading strategies from other firms. As the horizon of successful forecasting lengthens and the applicable market expands, the models could avoid damaging each others' performance.
Why this could meaningfully affect compute share through 2030
Frontier labs have orders of magnitude more compute than .15 GW today, even excluding prepaid commitments.[21] Part of the reason they have been able to capture so much compute share is simply the amount they are willing to pay for the capacity, which is driven by how much they can use it to earn. Across the labs, we see compute flowing to the model that can drive the highest margin. "You either die a frontier lab or live long enough to see yourself sell compute."[22] Right now, trading firms just don't have a revenue base to meaningfully drive up the cost of compute for the labs. Even if they can drive a higher margin on the compute they do purchase, the scale of that compute has been pretty small. Even if the entire industry started spending 40% of their revenues on compute, it would only pay for a gigawatt or two. But if buying compute directly drives increased trading revenue, the payback period is short enough, and the ceiling for alpha decay is high enough, compute spend from trading could get very big very fast.
especially for the trading firms, a reminder that these numbers are extremely speculative
It's time for a lot of naive regressions and lines on log plots! We'll just be looking at data from Jane Street and HRT, with lots of assumptions and estimates from what is public (mostly just compute commitments and total revenue by quarter). First, let's see if we can get a sense for the trend of revenue from compute intensive mid-frequency trading strategies. Depending on the exact share of trading revenue driven by MFT over time, it seems like it is fit by a trendline of about 32% growth per quarter (doubling every 7.6 months) since Q1 2024. We can back out an even rougher estimate for growth in compute commitments, which is growing around 35% per quarter. Assuming this holds through the end of the decade, it would grow the revenue base of the firms to over $2 trillion. This could support hundreds of billions in compute spend and would be comparable with the frontier labs' existing commitments. Furthermore, at this scale they would meaningfully drive up the cost of marginal compute for the labs (and for rogue/independent models).
part of the growth baked in here is an assumed growth in the share mid-frequency strategies in revenue, which is a completely unknown variable. this is super hand-wavy!
Notably, at 2 trillion the revenue of AI trading strategies would eclipse the current revenue of all trading organizations worldwide. These regressions are based on pretty weak evidence for such an extraordinary claim! However, for an intuition pump imagine a private model with a superhuman edge at forecasting global asset prices. Holding an edge of half a percent per year on global assets would be worth $2 trillion a year!
Compute is growing roughly 3x per year, which is currently outpaced by the extraordinary trend of 10x yearly growth of Anthropic's revenue over the past few years.[23] Dwarkesh implies that if this trend continues, compute prices must materially increase, unless the labs spend a larger share of compute on serving customers or their margins go much higher. I want to make the point that even if lab revenue does not increase on trend, the compute growth rate would be outpaced by the rate of trading compute growth alone (annualized at 3.3x/yr) and marginal compute costs may still increase. This has a few potential implications, if the trend does actually hold.
1. Trading may be the most profitable current application of transformer based models at scale.
Given the amount these trading firms are currently spending on compute compared to frontier labs, they are making a lot of money. Making a rough assumption about relative sources of trading revenues, compute is likely making them $5–$7 in profit for every $1 invested. Not only would this drive compute share their way, it would also incentivize the frontier labs to adopt a similar strategy. One of the answers to Dwarkesh's blog prize[24] presented this as a potential path to profitability, though I'd infer the labs would prefer avoiding it. However, they might not have a choice if compute markets are tempted by the potential returns trading presents compared to the speculative nature of AI research and model training.
2. The privatization of AI returns could remain into the far future.
I disagree with the "Open Global Investment" model of AI governance,[25] but the frontier labs going public does seem to offer some influence against pure private power centralization. If trading firms become a dominant power and source of both AI risk and return, their private corporate structure could keep the public from sharing in the returns without aggressive taxation.
3. If trading firms train large enough internal models, they could be a non-trivial source of catastrophic risk.
Most models of catastrophic misaligned AI risk (AI 2027, Adolescence of Technology, Situational Awareness)[26] make the assumption that powerful AIs will arise within governments or labs. However, as currently structured, there is some level of transparency innate in a business model that requires selling broad model access. On the government side, nobody has yet consolidated that much compute or research expertise. Within trading firms, models can be developed privately and their behavior can stay private while still generating the profit to fund future training runs. Commonly suggested AI regulation frameworks, such as embedded evaluators at frontier labs, would also entirely miss this risk vector. We have already seen 'misaligned' behavior from these trading organizations with humans at the helm. If trading responsibilities are handed over to models, especially under strict profit motives and spotty monitoring, we could see misaligned behavior with serious economic consequences.
Financial Times on Anthropic reaching profitability. ft.com↩︎
J.P. Morgan Global Research, "The AI-driven memory shortage" (jpmorgan.com); IDC, "Global memory shortage crisis: potential impact on the smartphone and PC markets in 2026" (idc.com). ↩︎
Galaxy, "Galaxy Completes Phase I of Its Helios Data Center Campus" (prnewswire.com); Core Scientific, "Core Scientific and CoreWeave Announce $1.2 Billion Expansion at Denton, TX Site" (corescientific.com). ↩︎
Jane Street, "Finding Signal in the Noise," Signals & Threads (signalsandthreads.com); "XTX Markets Launches New Machine Learning Division XTY Labs," LiquidityFinder (liquidityfinder.com). ↩︎
"XTX Markets Revenue Rises 43% to $3.93 Billion in 2025," Finance Magnates. financemagnates.com↩︎
"Jane Street Group Reports Record $20.5 Billion Net Trading Revenue," Bloomberg via Investing.com (investing.com); Business Insider on Hudson River Trading, syndicated copy (itiger.com). ↩︎
Business Insider on Hudson River Trading, March 2025, syndicated copy (itiger.com); Rupak Ghose, "The new Hudson River Trading," November 2025 (rupakghose.substack.com). ↩︎
Odd Lots, "How Hudson River Trading actually uses AI," October 31, 2025. podscan.fm↩︎
"Jane Street posts record $16.1 billion," Bloomberg via Yahoo Finance (finance.yahoo.com); "Hudson River Trading posts record $6.4 billion trading revenue in Q1," Bloomberg via Investing.com (investing.com). ↩︎
"Hudson River Trading racks up record $11.4bn revenue amid market volatility," Hedgeweek (hedgeweek.com); Rupak Ghose, "Jane Street is a hedge fund, stop" (rupakghose.substack.com); Fortune on Jane Street's July loss (fortune.com). ↩︎
CoreWeave 8-K exhibit, April 15, 2026 (sec.gov); "Lambda Partners with Hudson River Trading to Power Quantitative Research and Development" (businesswire.com); "Hudson River Trading expands Dell deployment to power AI-driven research" (dell.com). ↩︎
"Jane Street posts record $16.1 billion," Bloomberg via Yahoo Finance. finance.yahoo.com↩︎
"Flow Traders selects CoreWeave to power foundation model training for AI-driven quantitative trading" (coreweave.com); "IMC selects CoreWeave as firm deepens research investment" (coreweave.com); "QumulusAI Signs GPU-as-a-Service Agreement With DRW for NVIDIA Blackwell B300 Capacity" (businesswire.com). ↩︎
"Crusoe signs $13bn deal with Jane Street: report," Data Center Dynamics (datacenterdynamics.com); "Hudson River Trading to build next-gen research platform powered by NVIDIA Vera Rubin NVL72 on CoreWeave Cloud" (coreweave.com). ↩︎
Meng and Chen, "AI-Driven Alpha Decay: Algorithmic Homogenization, Reflexive Signal Erosion, and the Paradox of Intelligent Markets," March 2026. arxiv.org↩︎
Epoch AI, "How much AI compute do frontier labs use?" May 2026. epoch.ai↩︎
It's pretty crazy that right now, the highest margin thing to do with these models seems to be simply selling them through an API. Dwarkesh's blog prize [1] questioned how this dynamic could ever result in lab profitability, simply because the scale of reinvestment into model training and research requires constantly reinvesting more than you're making. Well, Anthropic is likely already profitable, [2] and it hasn't required any of the schemes I saw proposed in answers to his question. It turns out that the margin on selling frontier intelligence through an API is just really high! So high, that AI training and inference is consuming supply chains that used to serve other high-margin industries. Memory and fab space is being redirected from consumer devices, [3] and compute that used to serve Bitcoin mining is being repurposed to capture a slice of those frontier lab margins. [4]
This dynamic reinforces classic concentration of power risks, with some estimating OpenAI and Anthropic will soon control over 80% of worldwide compute. [5] As long as supplying frontier labs is the highest margin usage of compute, it's hard to not see this becoming the case. However, I think there is a decent case to be made that trading will become a threat to research compute, in the same way that research is now a threat to compute in consumer products, due to better profitability and the rapid growth.
Trading firm AI use
Hedge funds and trading firms have used "AI" for decades, as that amorphous label can apply to everything from algorithmic market making to high frequency trading. I want to focus on a specific trading strategy that seems to be gaining traction in that industry: compute intensive mid-frequency trading. As a reductive summary of the many different ways this is employed, this essentially involves training a model to predict the direction of a security over a short timescale, maybe 5 minutes, and trading on that assumption. The details of the specific trading strategies and data sources are closely guarded, and often target different asset classes or markets firm to firm (options, public equities, bonds, etc.). The compute needed to perform trading strategies has long been pretty meager. At the start of 2024, the most compute intensive player (Jane Street) had committed to only around $250 million in GPUs, and the MFT space as a whole was likely less than a billion. [6] These numbers don't touch the compute commitments of the labs and wouldn't be worth including in models of their economic power. That has changed.
A note on the numbers. Due to the secretive nature of trading, many of the numbers I'm throwing around are best estimates backed out of public announcements. These firms have incentives to both downplay and emphasize the success of certain strategies, which limits the ability to perform analysis like this. The people with the most visibility here are the firms themselves, so this is mostly based on their committed actions and disclosed returns. I also don't personally have a trading background! I come at this from an AI policy perspective, as I'm worried this risk might be overlooked.
In January of 2024, Hudson River Trading and Jane Street had around $600 million of contracted compute capacity combined. HRT's was mostly on-premises through an installed base of H100s, and Jane Street had just begun an agreement with CoreWeave to access a few thousand of their GPUs. This level of compute, even if paltry compared to the frontier labs, represented the lion's share of GPU capacity across Wall Street. The only exception might be XTX Markets, who had a significant amount of GPUs (25,000+!). However, they were focused on using it for specific low-latency market making strategies. [7] How HRT and Jane Street were using their compute was more unclear. In 2023, most of HRT's revenue had come from high frequency trading strategies and likely pulled in around $4 billion in revenue that year. Jane Street generated around $10 billion from a mix of approaches. In 2024, they both doubled revenue to nearly $8 billion and over $20 billion. [8]
These firms might simply have been cashing in on high levels of volatility. We can roughly compare their returns to other similar firms without compute. Citadel Securities' net revenue was about $6.3B in 2023 and $9.7B in 2024, increasing 55%. Virtu Financial had revenues of $2.29B in 2023 and $2.88B in 2024, up 24%. Optiver grew around 30%. [9] These firms are growing sizably and seem to be taking advantage of that sustained volatility, but none of them approached the growth seen by HRT or Jane Street that year.
Jane Street and HRT from bondholder disclosures reported by Bloomberg; Citadel Securities per Bloomberg; Virtu adjusted net trading income from SEC filings; Optiver from its annual results. 2026 is first-half revenue annualized (which honestly might undersell the trend) and for Jane Street Q2 is estimated. Lots of assumptions and napkin math, but the general shape should be correct.
Whatever caused this growth, both Jane Street and HRT started taking compute much more seriously after 2024. Jane Street announced expansions to tens of thousands of GPUs in the first half of 2025 and secondary sources reported that HRT was on track to spend a billion on compute that year. [10] Between the two of them, they ended up spending over $2 billion. No other firm in the space had disclosed any material compute spend. Reports came out that the companies were trying new kinds of trading. HRT in particular had mainly been known as a high-frequency trader and suddenly they were holding assets on their books longer to perform this mid-frequency trading. [11] On an Odd Lots podcast in late 2025, they directly attributed massive compute usage for the success of these kinds of strategies. [12] Again, both firms grew their trading revenues more than 50% on the year. Still, no other sizable competitor for trading compute had emerged.
In the first quarter of 2026, both firms pulled in more revenue than they had all of 2023. [13] Then the second quarter came in around double the first. [14] Both these firms began scrambling for compute; Jane Street signed a $6 billion commitment with CoreWeave in April, while HRT bought a billion of B200 capacity. [15] After Jane Street's $16B first quarter represented 40% of the previous year's revenue, heads started to turn on Wall Street. Previous anecdotes from firms like Citadel had dismissed many ML strategies to hold equities as "falling apart" when predicting further than 5 minutes into the future. [16] It sounded like Jane Street might be using models to recommend holding on the scale of days or even weeks. [17] Other firms followed their lead. Flow Traders, DRW, and IMC committed to purchasing a billion in compute capacity, from owning essentially none previously. [18] Jane Street and HRT had now both grown their revenues around 6x in the last two years, compared to less than 3x for any peers without meaningful compute.
By September, HRT and Jane Street already found themselves, again, compute constrained. Jane Street signed a deal valued at $13 billion over 5 years with Crusoe and HRT paid CoreWeave another few billion for capacity. [19] Of the around half a gigawatt of compute committed to trading, these two firms still control 80% of it. On the .15 GW in use this year, the firms could potentially generate over $150 billion in revenue. Again, exactly how much of this revenue is directly attributable to these strategies and how scalable they are with more compute is unclear, but the people who would know best are buying a hell of a lot more.
some of these deals are vague/multi-year but only show up in the quarter they were announced at the total contract value, which makes attribution and good modeling hard
One relevant piece of academic literature [20] suggests that this is not sustainable; that an increase in AI trading strategies will erode the alpha of all using them because of the correlation in the resulting trading behavior. This could very well happen, but a key difference with general AI trading strategies is that the models are not exploiting a specific market inefficiency. Rather, they would be superhuman at finding new inefficiencies to exploit. The models could, as well as the people writing that paper, understand how their signals would correlate with other AI trading strategies from other firms. As the horizon of successful forecasting lengthens and the applicable market expands, the models could avoid damaging each others' performance.
Why this could meaningfully affect compute share through 2030
Frontier labs have orders of magnitude more compute than .15 GW today, even excluding prepaid commitments. [21] Part of the reason they have been able to capture so much compute share is simply the amount they are willing to pay for the capacity, which is driven by how much they can use it to earn. Across the labs, we see compute flowing to the model that can drive the highest margin. "You either die a frontier lab or live long enough to see yourself sell compute." [22] Right now, trading firms just don't have a revenue base to meaningfully drive up the cost of compute for the labs. Even if they can drive a higher margin on the compute they do purchase, the scale of that compute has been pretty small. Even if the entire industry started spending 40% of their revenues on compute, it would only pay for a gigawatt or two. But if buying compute directly drives increased trading revenue, the payback period is short enough, and the ceiling for alpha decay is high enough, compute spend from trading could get very big very fast.
especially for the trading firms, a reminder that these numbers are extremely speculative
It's time for a lot of naive regressions and lines on log plots! We'll just be looking at data from Jane Street and HRT, with lots of assumptions and estimates from what is public (mostly just compute commitments and total revenue by quarter). First, let's see if we can get a sense for the trend of revenue from compute intensive mid-frequency trading strategies. Depending on the exact share of trading revenue driven by MFT over time, it seems like it is fit by a trendline of about 32% growth per quarter (doubling every 7.6 months) since Q1 2024. We can back out an even rougher estimate for growth in compute commitments, which is growing around 35% per quarter. Assuming this holds through the end of the decade, it would grow the revenue base of the firms to over $2 trillion. This could support hundreds of billions in compute spend and would be comparable with the frontier labs' existing commitments. Furthermore, at this scale they would meaningfully drive up the cost of marginal compute for the labs (and for rogue/independent models).
part of the growth baked in here is an assumed growth in the share mid-frequency strategies in revenue, which is a completely unknown variable. this is super hand-wavy!
Notably, at 2 trillion the revenue of AI trading strategies would eclipse the current revenue of all trading organizations worldwide. These regressions are based on pretty weak evidence for such an extraordinary claim! However, for an intuition pump imagine a private model with a superhuman edge at forecasting global asset prices. Holding an edge of half a percent per year on global assets would be worth $2 trillion a year!
Compute is growing roughly 3x per year, which is currently outpaced by the extraordinary trend of 10x yearly growth of Anthropic's revenue over the past few years. [23] Dwarkesh implies that if this trend continues, compute prices must materially increase, unless the labs spend a larger share of compute on serving customers or their margins go much higher. I want to make the point that even if lab revenue does not increase on trend, the compute growth rate would be outpaced by the rate of trading compute growth alone (annualized at 3.3x/yr) and marginal compute costs may still increase. This has a few potential implications, if the trend does actually hold.
1. Trading may be the most profitable current application of transformer based models at scale.
Given the amount these trading firms are currently spending on compute compared to frontier labs, they are making a lot of money. Making a rough assumption about relative sources of trading revenues, compute is likely making them $5–$7 in profit for every $1 invested. Not only would this drive compute share their way, it would also incentivize the frontier labs to adopt a similar strategy. One of the answers to Dwarkesh's blog prize [24] presented this as a potential path to profitability, though I'd infer the labs would prefer avoiding it. However, they might not have a choice if compute markets are tempted by the potential returns trading presents compared to the speculative nature of AI research and model training.
2. The privatization of AI returns could remain into the far future.
I disagree with the "Open Global Investment" model of AI governance, [25] but the frontier labs going public does seem to offer some influence against pure private power centralization. If trading firms become a dominant power and source of both AI risk and return, their private corporate structure could keep the public from sharing in the returns without aggressive taxation.
3. If trading firms train large enough internal models, they could be a non-trivial source of catastrophic risk.
Most models of catastrophic misaligned AI risk (AI 2027, Adolescence of Technology, Situational Awareness) [26] make the assumption that powerful AIs will arise within governments or labs. However, as currently structured, there is some level of transparency innate in a business model that requires selling broad model access. On the government side, nobody has yet consolidated that much compute or research expertise. Within trading firms, models can be developed privately and their behavior can stay private while still generating the profit to fund future training runs. Commonly suggested AI regulation frameworks, such as embedded evaluators at frontier labs, would also entirely miss this risk vector. We have already seen 'misaligned' behavior from these trading organizations with humans at the helm. If trading responsibilities are handed over to models, especially under strict profit motives and spotty monitoring, we could see misaligned behavior with serious economic consequences.
Dwarkesh Patel, "Blog prize." dwarkesh.com ↩︎
Financial Times on Anthropic reaching profitability. ft.com ↩︎
J.P. Morgan Global Research, "The AI-driven memory shortage" (jpmorgan.com); IDC, "Global memory shortage crisis: potential impact on the smartphone and PC markets in 2026" (idc.com). ↩︎
Galaxy, "Galaxy Completes Phase I of Its Helios Data Center Campus" (prnewswire.com); Core Scientific, "Core Scientific and CoreWeave Announce $1.2 Billion Expansion at Denton, TX Site" (corescientific.com). ↩︎
Dwarkesh Patel interviewing Dylan Patel. dwarkesh.com ↩︎
Jane Street, "Finding Signal in the Noise," Signals & Threads (signalsandthreads.com); "XTX Markets Launches New Machine Learning Division XTY Labs," LiquidityFinder (liquidityfinder.com). ↩︎
"XTX Markets Revenue Rises 43% to $3.93 Billion in 2025," Finance Magnates. financemagnates.com ↩︎
"Jane Street Group Reports Record $20.5 Billion Net Trading Revenue," Bloomberg via Investing.com (investing.com); Business Insider on Hudson River Trading, syndicated copy (itiger.com). ↩︎
"Griffin's Citadel Securities reports record $9.7bn trading revenue," Hedgeweek (hedgeweek.com); Virtu Financial fourth-quarter 2024 results (sec.gov); "Optiver reports strong financial results for 2024" (optiver.com). ↩︎
"Jane Street compute," eFinancialCareers (efinancialcareers-gulf.com); Jane Street, "A look inside our newest data center" (janestreet.com); "How Hudson River Trading is breaking records in 2025," Disruption Banking (disruptionbanking.com). ↩︎
Business Insider on Hudson River Trading, March 2025, syndicated copy (itiger.com); Rupak Ghose, "The new Hudson River Trading," November 2025 (rupakghose.substack.com). ↩︎
Odd Lots, "How Hudson River Trading actually uses AI," October 31, 2025. podscan.fm ↩︎
"Jane Street posts record $16.1 billion," Bloomberg via Yahoo Finance (finance.yahoo.com); "Hudson River Trading posts record $6.4 billion trading revenue in Q1," Bloomberg via Investing.com (investing.com). ↩︎
"Hudson River Trading racks up record $11.4bn revenue amid market volatility," Hedgeweek (hedgeweek.com); Rupak Ghose, "Jane Street is a hedge fund, stop" (rupakghose.substack.com); Fortune on Jane Street's July loss (fortune.com). ↩︎
CoreWeave 8-K exhibit, April 15, 2026 (sec.gov); "Lambda Partners with Hudson River Trading to Power Quantitative Research and Development" (businesswire.com); "Hudson River Trading expands Dell deployment to power AI-driven research" (dell.com). ↩︎
Ken Griffin at Stanford GSB, May 2025. aol.com ↩︎
"Jane Street posts record $16.1 billion," Bloomberg via Yahoo Finance. finance.yahoo.com ↩︎
"Flow Traders selects CoreWeave to power foundation model training for AI-driven quantitative trading" (coreweave.com); "IMC selects CoreWeave as firm deepens research investment" (coreweave.com); "QumulusAI Signs GPU-as-a-Service Agreement With DRW for NVIDIA Blackwell B300 Capacity" (businesswire.com). ↩︎
"Crusoe signs $13bn deal with Jane Street: report," Data Center Dynamics (datacenterdynamics.com); "Hudson River Trading to build next-gen research platform powered by NVIDIA Vera Rubin NVL72 on CoreWeave Cloud" (coreweave.com). ↩︎
Meng and Chen, "AI-Driven Alpha Decay: Algorithmic Homogenization, Reflexive Signal Erosion, and the Paradox of Intelligent Markets," March 2026. arxiv.org ↩︎
Epoch AI, "How much AI compute do frontier labs use?" May 2026. epoch.ai ↩︎
roon (@tszzl) on X. x.com ↩︎
Dwarkesh Patel, "Why compute might get 10x more expensive." dwarkesh.com ↩︎
"How the AI labs make profit (maybe, eventually)," LessWrong. lesswrong.com ↩︎
Nick Bostrom, "Open Global Investment as a Governance Model for AGI," 2025. nickbostrom.com ↩︎
AI 2027; Dario Amodei, The Adolescence of Technology; Leopold Aschenbrenner, Situational Awareness. ↩︎