H200 rent: ~+160% since January 1.
B200: ~+120%. H100: ~+35%.
Even the six-year-old A100 rents for more than it did in January.
Tell us the GPU type, quantity and duration, and we'll quote.
Booking early, or waiting it out?
a16z.news/p/charts-of-th…
B200: ~+120%. H100: ~+35%.
Even the six-year-old A100 rents for more than it did in January.
Tell us the GPU type, quantity and duration, and we'll quote.
Booking early, or waiting it out?
a16z.news/p/charts-of-th…
0
0
0
24
Flapping Airplanes raised a $1B Series A, co-led by Index and Lux, on one bet: more useful work from every GPU.
Builders can make it too: match the card to the job, keep it busy, give it back when done.
RTX 4090 from $0.49/hr. Send GPU model, quantity and duration for a quote.
Builders can make it too: match the card to the job, keep it busy, give it back when done.
RTX 4090 from $0.49/hr. Send GPU model, quantity and duration for a quote.
0
0
0
34
GPU rental prices went up again this week, per Silicon Data:
H100 +10.3% to $2.88/hr
H200 +6.0% to $3.51/hr
B200 +9.0% to $6.29/hr
B300 +5.8% to $7.35/hr
Waiting for prices to fall has been a losing bet all year.
Tell us GPU type, quantity and duration — we'll send a quote.
H100 +10.3% to $2.88/hr
H200 +6.0% to $3.51/hr
B200 +9.0% to $6.29/hr
B300 +5.8% to $7.35/hr
Waiting for prices to fall has been a losing bet all year.
Tell us GPU type, quantity and duration — we'll send a quote.
0
0
0
45
GPU rental prices aren't going down.
IREN's co-CEO on CNBC: 2026 is largely sold out; new 3-year deals at more than 2x last year.
That's for people buying years. Most teams need days.
4090 from $0.49/hr, H100 from $2.96/hr, billed per second. Tell us GPU, how many, how long.
IREN's co-CEO on CNBC: 2026 is largely sold out; new 3-year deals at more than 2x last year.
That's for people buying years. Most teams need days.
4090 from $0.49/hr, H100 from $2.96/hr, billed per second. Tell us GPU, how many, how long.
0
0
1
42
When demand is "off the charts," big buyers book whole data centers. The team that needs 2 GPUs for a week gets squeezed.
Nebius just lined up $3.5B from Citi to build more. Capacity is coming.
The question is who gets it. Glows is for that small team.
Nebius just lined up $3.5B from Citi to build more. Capacity is coming.
The question is who gets it. Glows is for that small team.
Nvidia CEO Jensen Huang says neocloud demand is 'off the charts' but capital is scarce cnbc.com/2026/10/22/nvi…
0
0
0
33
One researcher, agents writing 99%+ of the code, 3.5 months, under $10M of compute all-in: that's how Mirendil says it pretrained Mirendil-Base.
The bottleneck is moving from headcount to GPU-hours.
H100 80GB from $2.96/hr. Send GPU model, quantity and duration for a quote.
The bottleneck is moving from headcount to GPU-hours.
H100 80GB from $2.96/hr. Send GPU model, quantity and duration for a quote.
0
0
1
21
IREN's planned ~300MW Spain campus targets 2028. NVIDIA plans to ship DSX POD from early 2027.
This quarter needs GPUs up now:
RTX 4090 24GB from $0.49/hr
L40S 48GB from $0.83/hr
A100 80GB from $1.20/hr
H100 80GB from $2.96/hr
Send GPU model, quantity and duration for a quote.
This quarter needs GPUs up now:
RTX 4090 24GB from $0.49/hr
L40S 48GB from $0.83/hr
A100 80GB from $1.20/hr
H100 80GB from $2.96/hr
Send GPU model, quantity and duration for a quote.
0
0
1
40
Spot or on-demand? Wrong question. Ask what a stop costs you.
Checkpointed batch: resumes. Spot.
Fine-tune with frequent saves: a few steps. Spot.
Long run, no checkpoints: everything. On-demand.
Serving or interactive dev: on-demand.
Price the restart, not just the hour.
Checkpointed batch: resumes. Spot.
Fine-tune with frequent saves: a few steps. Spot.
Long run, no checkpoints: everything. On-demand.
Serving or interactive dev: on-demand.
Price the restart, not just the hour.
0
0
0
24
ByteDance, Alibaba and Tencent reportedly ordered 3M+ RTX PRO 5500 workstation cards, mostly for inference (Reuters).
Inference needs VRAM for weights and KV cache, not always a flagship.
RTX 6000 Ada 48GB from $0.72/hr, RTX PRO 6000 96GB from $1.68/hr. Send GPU model, quantity e.
Inference needs VRAM for weights and KV cache, not always a flagship.
RTX 6000 Ada 48GB from $0.72/hr, RTX PRO 6000 96GB from $1.68/hr. Send GPU model, quantity e.
0
0
0
46
"2026 is largely sold out," and recent 3-year contracts are at more than double the prior year's pricing, per IREN's CEO.
That's what a seller's market looks like.
Multi-year take-or-pay makes sense for a hyperscale build-out. It doesn't fit a 6-hour fine-tune or a weekend eval terms. Glows on-demand GPUs are pay-as-you-go, billed per second with no hourly minimum: RTX 4090 from $0.49/hr, L40S from $0.83/hr, H100 from $2.96/hr.
Tell us the GPU, quantity and duration, and we'll quote it.
That's what a seller's market looks like.
Multi-year take-or-pay makes sense for a hyperscale build-out. It doesn't fit a 6-hour fine-tune or a weekend eval terms. Glows on-demand GPUs are pay-as-you-go, billed per second with no hourly minimum: RTX 4090 from $0.49/hr, L40S from $0.83/hr, H100 from $2.96/hr.
Tell us the GPU, quantity and duration, and we'll quote it.
A lot of questions this week on funding, so here's how a GPU typically gets funded at $IREN.
First, the customer signs a multi-year take-or-pay contract and agrees to a prepayment. On our recent 3-year contracts, prepayments have been 45% to 55% of the associated GPU capex.
Second, financing is committed, secured by the contracted GPUs. On the Microsoft contract we secured an A-rated facility at around 6%, which together with prepayments covers 96% of GPU capex. More recently, for a non-investment grade counterparty at Mackenzie, we secured financing at 9% fixed, which together with prepayments covers more than the full cost of the GPUs.
That's the GPU side. The data centers are a separate, long-term asset base, and we own the ones our GPUs run in. Construction has been funded from cash to date, with no debt against these multi-billion dollar assets. As they come online with contracted customers, they can support financing in their own right.
The processes behind the next facilities are continuing, and new paper will cost what the market charges. With GPU paybacks of around two years on our recent 3-year contracts, a wider spread barely moves the economics.
There's more customer demand than we can build for. As we noted at our recent earnings, 2026 is largely sold out, recent 3-year contracts are at more than double the prior year's pricing, and demand for anything we can bring online in 2027 covers it several times over, with prepayments a focus in every negotiation. Conversations on 2028 capacity are already underway.
Meanwhile announced capacity keeps slipping across the industry, which is the point: signing deals isn't the bottleneck in this market, bringing GPUs online is, and that's our focus.
1
0
1
30
DeepSeek open-sourced DeepInfer, the engine behind V4.1-Flash. Read the hardware line before you clone it.
Per the README:
→ One model: DeepSeek-V4.1-Flash
→ NVIDIA SM100 GPUs or Huawei Ascend 950
→ CUDA 13, Python 3.12
→ Benchmarked on 4 GPUs for prefill, 64 for decode
→ cache system, prefill/decode disaggregation, load balancing across replicas
Open engine. Not a turnkey service.
Size the workload first: model, context, traffic. Then tell us the GPU, quantity and duration, and we'll quote it.
Per the README:
→ One model: DeepSeek-V4.1-Flash
→ NVIDIA SM100 GPUs or Huawei Ascend 950
→ CUDA 13, Python 3.12
→ Benchmarked on 4 GPUs for prefill, 64 for decode
→ cache system, prefill/decode disaggregation, load balancing across replicas
Open engine. Not a turnkey service.
Size the workload first: model, context, traffic. Then tell us the GPU, quantity and duration, and we'll quote it.
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
1
0
0
47
We ran four Qwen3.5 models on one 48GB L40S: 4B, 9B, 27B and 35B-A3B.
The bigger dense ones drew a better pelican but slowed from 102 to 21 tok/s.
The 35B-A3B is a MoE. It drew the best pelican and was the fastest of all four at 122 tok/s.
What would you deploy on 48GB VRAM card?
The bigger dense ones drew a better pelican but slowed from 102 to 21 tok/s.
The 35B-A3B is a MoE. It drew the best pelican and was the fastest of all four at 122 tok/s.
What would you deploy on 48GB VRAM card?
1
0
2
179
The bottleneck for AI agents isn't capability anymore. It's design.
At AIE New York this week, OpenAI's Joseph Kim, who leads ChatGPT for financial services, called it the threshold of usefulness: the point where a model is reliable enough to actually save you time.
Below it, tizes mistakes. Above it, handing work to an agent finally pays off.
His personal bet: by 2027, finance leans far less on Excel and PowerPoint, and agents take on idea generation and trading tasks.
What gets an agent over that line isn't a bigger model. It's three design calls:
→ Where is the threshold for this task?
→ Where do humans review: before, during, or after?
→ Can every number be traced back to its source?
More guardrails don't make agents safe. Control points in the right places do.
None of this waits for the next model. Builders can start designing it now.
Build AI here.
At AIE New York this week, OpenAI's Joseph Kim, who leads ChatGPT for financial services, called it the threshold of usefulness: the point where a model is reliable enough to actually save you time.
Below it, tizes mistakes. Above it, handing work to an agent finally pays off.
His personal bet: by 2027, finance leans far less on Excel and PowerPoint, and agents take on idea generation and trading tasks.
What gets an agent over that line isn't a bigger model. It's three design calls:
→ Where is the threshold for this task?
→ Where do humans review: before, during, or after?
→ Can every number be traced back to its source?
More guardrails don't make agents safe. Control points in the right places do.
None of this waits for the next model. Builders can start designing it now.
Build AI here.
1
0
3
65