Tag
OpenAI now lets one API key target us., eu. or in.api.openai.com per request. Only the US, EU and UAE support regional inference. India, the UK and Japan are storage-only.
Manu Shukla 
Three IEEE S&P 2026 papers turned GPU Rowhammer into a privilege-escalation problem. GPUBreach reached a root shell with the IOMMU on. Which cards flip, what ECC costs, what to change now.
Manu Shukla 
Rapid Bucket is 5.5x the per-GiB price of standard Cloud Storage and bills for reads a same-region bucket serves free. On 50 TiB of checkpoints that is about $5,243 a month extra. Here is the arithmetic.
Manu Shukla 
In 2026 an AI-driven memory shortage pushed DDR5 and server DRAM to record prices and stretched lead times. Here is what changed, why, and a sourced framework for CTOs choosing to buy hardware now, shift to cloud
Manu Shukla 
In 2026, 98% of FinOps teams manage AI spend, and LLM prices span more than 140x from budget to frontier. An AI gateway with cost-aware routing and caching cuts that bill 40-70% without downgrading quality. How eCorpIT
Manu Shukla 
Kimi K3's open weights are 594 GB in BF16 and need about 1.68 TB of VRAM to serve; the API lists at $3/$15 per 1M tokens. Here is the 2026 self-host vs API break-even math with real GPU rates.
Manu Shukla 
AMD's Instinct MI400 series splits in two: the MI455X for frontier AI and inference, the MI430X for sovereign AI and FP64 HPC. A decision guide to the specs, the cost claims and 2026-2027 availability.
Manu Shukla 
AMD's ROCm 7, plus SCALE and HIPIFY, make some CUDA workloads portable to AMD GPUs. But custom kernels, TensorRT-LLM and NCCL still bind enterprises to NVIDIA. What actually moves off CUDA in 2026.
Manu Shukla 
A specs-and-cost comparison of Google Ironwood (TPU7x) and NVIDIA B200 for AI inference in 2026: memory, bandwidth, scale, software lock-in, consumption models, and a decision matrix for when to pick each.
Manu Shukla 
India's data centre numbers get quoted interchangeably, but 1.9 GW of capacity, $120 billion in commitments, and a $1.7 billion market are three different things. Here is how to read them before you decide where to host
Manu Shukla 
AMD and NVIDIA both ship 72-accelerator AI racks in 2026, and almost none of their headline numbers are measured the same way. A line-by-line normalisation of the two vendor-published spec sheets, and the seven
Manu Shukla 
AWS made Amazon Bedrock Managed Knowledge Base generally available on 17 June 2026. It removes the vector store, the parsing pipeline and the retrieval orchestration. It also removes some control, and it is not in
Manu Shukla 
Karnataka said on 15 July 2026 that a new data centre policy is coming. A 2022-2027 policy is already published, and the change that actually moved the economics this year came from the Union Budget.
Manu Shukla 
Gartner expects power to constrain 40% of AI data centers by 2027, and GPU lead times run 36 to 52 weeks. This 2026 playbook covers capacity forecasting, spot vs reserved, neo-clouds and procurement governance.
Manu Shukla 
Meta reportedly plans Meta Compute to rent out its spare AI GPUs, and the news moved billions on 1 July 2026. What is confirmed, why the neoclouds fell hardest, and whether renting surplus GPU capacity is worth it.
Manu Shukla 
A senior-engineer comparison of RAG embedding models in 2026: NVIDIA Nemotron 3 Embed, OpenAI text-embedding-3, Cohere Embed v4 and open-weight Qwen3, with RTEB benchmarks, pricing and self-host break-even math.
Manu Shukla 
Kubernetes v1.35 introduced the Workload API and native gang scheduling. Four months later v1.36 replaced the API version entirely. Here is the current shape, the migration, and the GPU maths that justifies it.
Manu Shukla 
AI infrastructure spending is projected at $1.5 trillion for 2026, and Sequoia's arithmetic says the industry must earn $3 trillion to justify it. Frontier lab revenue sits in the low hundreds of billions. Here is what
Manu Shukla 
AWS lifted EC2 Capacity Block rates about 20% in July 2026 and 15% more on October 7. P6-B300 now bills at $16.146 per accelerator hour. What changed, why analysts disagree on the cause
Manu Shukla 
vLLM's April 2026 benchmark put FP8 KV-cache 14.9% above BF16 on one H100. Priced out, that is $4.24 down to $3.69 per million output tokens - and your AWS purchase model moves the number four times further.
Manu Shukla 
Kubernetes is the default home for AI, yet GPUs often run at 5% utilization, wasting over $100,000 a year. How a managed platform fixes GPU scheduling and cost for Indian enterprises running AI.
Manu Shukla 
Bloomberg reported on July 1, 2026 that Meta is building a cloud business to rent out AI compute and host Llama models. A fourth serious cloud vendor changes the math on price, negotiation, and lock-in for enterprise
Manu Shukla 
NVIDIA's Vera Rubin began production shipments in August 2026 and is reaching cloud partners now, promising 5x inference over Blackwell and up to 10x lower cost per token. What to plan for.
Manu Shukla 
At WWDC 2026 Apple extended Private Cloud Compute to Google Cloud on Nvidia Blackwell GPUs. Here are five architecture decisions CTOs should take from Apple's private-AI design.
Manu Shukla 