40k?
Easy, spend 15k in Anthropic highest subscription & the rest in API
The rest? Spend it on a H100 (27k), then dump the rest into hardware

I think you don't need to spend money to get $40k in compute power, but it just takes longer. But you probably already have more hardware than me.

But i'd give some of my spare compute power overnight, but i wouldn't expect it to be all that 'fast'.

rent a b200 x8 on Google Cloud.

Load GLM-4.7-Flash and generate 15 Billion Rows.

according to chatgpt If by “15 billion rows” you mean generating a giant synthetic dataset with an LLM — like JSON rows, chat records, code samples, embeddings metadata, etc. — then the bottleneck is almost entirely tokens/sec, not raw FLOPS. And an 8× NVIDIA B200 cluster is an absolute monster for inference 😵‍💫⚡
Here’s the practical estimate.

Assumptions 📊
I’ll assume:

Model: “GLM-4.7-Flash” sized roughly like a fast 30B–70B inference model

Hardware: 8× B200 HGX node

Engine: vLLM / TensorRT-LLM

Quantization: FP8 or INT8

Each row:

~100 output tokens average

lightweight structured text

So:
15 billion rows × 100 tokens= 1.5 trillion tokens generated

Realistic Throughput
Modern B200 clusters are hitting absurd inference rates.
A public benchmark showed ~1.1 million tokens/sec on 96 B200s for a 27B model.
Scaling that down:
1.1M tok/s ÷ 96 GPUs≈ 11.5k tok/s per GPU× 8 B200s≈ 92k tok/s
That’s conservative. With batching and optimized serving you might see:

80k–150k tok/s sustained

Time Estimate ⏱️
Using the 1.5 trillion token assumption:
Conservative
1.5T ÷ 80k tok/s≈ 18.75 million seconds≈ 217 days
Optimized
1.5T ÷ 150k tok/s≈ 10 million seconds≈ 116 days
So:
Final Answer
Generating 15 billion rows with a large GLM model on an 8× B200 node would likely take:
👉 ~4 to 7 months continuously
…assuming:

~100 tokens per row

one single 8-GPU node

high utilization

no interruptions

But Here’s The Twist 🌪️
The row size changes everything.
Avg Tokens/RowTotal TokensEstimated Time10150B11–22 days25375B1–2 months1001.5T4–7 months5007.5T2–3 years
That’s why hyperscalers obsess over token efficiency. Tiny schema changes become months of GPU time.

Cost Reality 💸
An 8× B200 cloud instance is likely:

~$40–$100/hour depending on provider and reservation

So 4–7 months nonstop becomes roughly:
$120k – $500k+
for a single run 😭
At this scale, companies usually:

shard across many nodes

use smaller distilled models

generate compressed templates

use speculative decoding

generate structured latent data instead of raw prose

Because brute-forcing trillions of tokens is basically building a mini synthetic-data factory.
The silicon becomes a furnace. The datacenter becomes weather. 🌩️🖥️

My brother, I love you—you are simply magnificent! Is the line between genius and madness truly just a single step? You are as mad as a genius scientist like Nikola Tesla. I am, however, a bit concerned by your behavior; you actually went so far as to mortgage your own home for the sake of scientific advancement. You are truly extraordinary. I love you, my brother, my friend.

My brother, I love you—you are simply magnificent! Is the line between genius and madness truly just a single step? You are as mad as a genius scientist like Nikola Tesla. I am, however, a bit concerned by your behavior; you actually went so far as to mortgage your own home for the sake of scientific advancement. You are truly extraordinary. I love you, my brother, my friend.