
The Capital Markets Layer for AI Compute. Making compute fungible and tradable. Backed by @HashKey_Capital & @FBGCapital
Global
Joined December 2024
- Tweets1.2K
- Following259
- Followers73.8K
- Likes1.7K
Pinned Tweet
We’re making compute fungible and tradable.
Watch how it happens and join the waitlist 👇
app.rax.finance/waitlist/
That’s what RAX Finance builds: one standard unit, the
compute token (cT).
1 cT = 1 GPU-hour, backed by a real machine
1
3
258
Once compute is priced by region, hedging, settlement
and financing can be built on top.
Join the waitlist: app.rax.finance/waitlist/
3
223
🔹RAX Finance · Weekly Update #01🔹
Locked in. Building.
A quieter week on the timeline, a busy one behind the scenes.
Starting today, we’re sharing a Weekly Update every Friday.
Stay tuned 🟢
Compute has an expiration problem
An unused GPU-hour can't be stored for tomorrow.
Once that hour passes, the capacity is gone
But access to compute can be standardized, priced,
and traded before it's consumed.
That's what we're building at RAX Finance
Making compute fungible and tradable.
🤖 Made with AI
🚀 RAX Ambassador Round 2 is OPEN!
🎯 10 slots available
📅 Apply within 5 days
♾️ Ongoing program (3-strike inactivity rule)
Weekly Missions
→ Content Creation: +10–30 pts
→ Community Support: +5 pts
→ Up to 35 pts/week
🎁 Weekly Unboxing Opportunity rewards for ambassadors.
Your points contribute to your reward weight.
Who we’re looking for:
✓ Consistent contributors
✓ Community-minded participants
✓ Long-term thinkers
Not just here for a one-off post — we value people who show up and grow with us.
📝 Apply now:
forms.gle/822M3vSjTxwU6JZz7
Be the Architect. Own the Engine.
You can buy a barrel of oil without knowing which well
it came from.
But renting GPU compute still means choosing a
provider, region, hardware configuration, and contract.
That's the difference between buying a commodity
and negotiating a service.
AI compute is ready for a better market structure.
And that's exactly what RAX Finance is building.
rax.finance
🤖 Made with AI
643 AI data center campuses. Across 55 countries.
That's the infrastructure footprint captured in one
recent industry census.
And every new location brings its own energy costs,
capacity constraints, and compute economics.
AI may be global, but the cost of powering it isn't
universal.
🤖 Made with AI
A GPU doesn't generate revenue just because it's
powerful
It generates revenue when someone needs its
compute.
That distinction matters a lot more than people think.
🤖 Made with AI
Scaling AI isn’t just about getting more GPUs.
what do you think is the most underestimated
bottleneck?
39%Power
13%Hardware
24%Networking
24%Economics
46 votes • Final resultsThere's a difference between having a price and having a market.
GPU providers already publish prices. But those quotes vary by region, hardware, contract length, and availability.
A functioning commodity market needs something more: standardized units, reliable delivery, and prices participants can actually trade against.
That's where compute markets get interesting.
🤖 Made with AI
What is an hour of AI compute actually worth?
Not all compute is priced the same.
The value of AI compute changes based on:
• GPU model
• Location
• Availability
• Demand
As AI infrastructure scales, transparent price discovery becomes essential.
That’s the market infrastructure RAX is building.
Hot take time:
AI compute should be priced like:
A) Oil — global commodity, single benchmark (WTI/Brent model)
B) Electricity — regional grid, locational pricing (LMP model)
Reply with A or B + your reasoning.
We're clearly Team B (hence building regional indices), but curious what you think.
Especially want to hear from folks who disagree 👀
Friday check-in 👋
What’s caught your attention in AI compute this week?
A new GPU, a data centre trend, or the future of ComputeFi?
We’re curious what the Rax community is following. Drop your thoughts below 👀
📊 Open question for the compute market:
By the end of 2027, what % of APAC third-party GPU capacity will be traded on secondary markets rather than remain locked in long-term contracts?
We’re talking about neoclouds and independent operators, excluding hyperscaler self-use. Secondary trading includes unused or residual contracted capacity changing hands.
Our working hypothesis: 15–25%.
Why?
• Multi-year take-or-pay contracts still dominate, with limited early-exit options.
• Secondary trading is emerging, but liquidity, standardisation, and counterparty trust are still developing.
• As market infrastructure matures, more unused capacity could find a second buyer.
What’s your prediction, and why? 👀
Let’s revisit this in 12 months and see who got closest.
Compute is being planned on a multi-year horizon.
Chipmakers are already securing EUV tools for 2029 and 2030.
ASML says customers are asking for more machines every day, even as 2027–2028 capacity expands.
The industry is locking in the manufacturing base for the next wave of AI capacity years in advance.
Via @wallstengine
$ASML SAYS CHIPMAKERS ARE ALREADY LOOKING TO PLACE EQUIPMENT ORDERS FOR 2029 AND EVEN 2030
CEO Christophe Fouquet says “customers are asking us every single day to get more tools,” as AI-related investments drive stronger demand for advanced semiconductor equipment.
The company also outlined major milestones for its next-generation High-NA EUV technology, which Fouquet says is finally approaching high-volume manufacturing after more than a decade of development.
Samsung plans to begin using High-NA EUV in mass production in 2028, followed by TSMC around 2030. Intel is already increasing the number of layers processed and wafer output using the technology.
Separately, ASML plans to introduce its NXE:3800F EUV system in 2027, increasing throughput from 230 to 260 wafers per hour, with a greater impact on customer manufacturing capacity expected from 2028 onward.













