Sameer Koli

Geek | Tech Enthusiast | Software Professional | Opinionated | 😀

[Sameer Koli @sameergkoli]

DeepSeek just graduated V4 to stable — the real change is peak-hour pricing: 2x baseline during Beijing business hours. If your inference cost model ignores time-of-day pricing on open-weight APIs, you're about to find it in next month's bill.

[Sameer Koli @sameergkoli]

OpenAI just raised its 2030 compute budget from $600B to $750B. The real signal: a $20B Georgia data center it's building itself, not leasing. Once your AI vendor becomes your landlord, your pricing leverage is gone for a decade — not a contract cycle.

[Sameer Koli @sameergkoli]

I just published Multi-Agent AI Skipped the One Microservices Lesson That Actually Mattered https://t.co/ls9qQM3SN2

[Sameer Koli @sameergkoli]

Google delayed Gemini 3.5 Pro to July while four senior DeepMind researchers left for Anthropic — and the market wiped $225B off Alphabet. Lesson for product leaders: a slipped date in AI isn't a scheduling problem, it's a capability signal. The market prices your bench first.

[Sameer Koli @sameergkoli]

Fireworks AI raised $1.5B at $17.5B — not to build a better model, but to run other people's models cheaper. 40T tokens/day, $1B ARR, Nvidia on the cap table. As models commoditize, margin shifts to the serving layer — GPU utilization, routing, queue depth, not weights.

[Sameer Koli @sameergkoli]

Microsoft just became the first hyperscaler to commit "at scale" to a non-Nvidia rack for production inference. Not because AMD is faster — because no CTO wants one vendor's roadmap dictating their unit economics for five years. Watch enterprise AI teams do the same next.

[Sameer Koli @sameergkoli]

I just published MCP Won the Standards War. It Didn’t Win the Security Argument. https://t.co/5L480INKWF

[Sameer Koli @sameergkoli]

Lenders wouldn't touch inference chips as collateral until Upper90's $400M General Compute loan this week. That's the real signal: inference capacity now holds resale value on its own — the same collateral test CTPOs should run before any buy-vs-rent GPU decision.

[Sameer Koli @sameergkoli]

Every regulated-industry AI deployment I've reviewed has the same hole: nobody named who owns an agent's decision before it shipped. Governance tooling gets added after the incident, not before the rollout. That's not a compliance gap. It's an unassigned accountability line.

[Sameer Koli @sameergkoli]

I just published Your GPU Bill Isn’t a Capacity Problem. It’s a Utilization Problem. https://t.co/PFIU59kGBh

[Sameer Koli @sameergkoli]

China just formed a 29-country AI governance body, WAICO. Brazil, Pakistan, Russia are founding members. India isn't, and the window closes July 30. Skipping a rulemaking table doesn't make you neutral, it just means you'll live by rules you didn't help write.

[Sameer Koli @sameergkoli]

RBI's draft "no black-box AI" mandate isn't a brake on adoption — it's a vendor filter. Most GenAI stacks sold to Indian BFSI can't explain a denied transaction, let alone kill-switch mid-inference. Explainability-first teams just got a moat everyone else must retrofit.

[Sameer Koli @sameergkoli]

I just published Your GPU Bill Isn’t a Capacity Problem. It’s a Utilization Problem. https://t.co/tenXg5T9db

[Sameer Koli @sameergkoli]

The UK just put AWS, Google Cloud, Microsoft and Oracle under systemic-risk supervision — the tier for banks too big to fail. If your BFSI stack runs on one hyperscaler, that's no longer a compliance footnote, it's a regulator's headline. Multi-cloud is now a board conversation.

[Sameer Koli @sameergkoli]

StanChart just cut over 7,000 roles and called it what it is: AI replacing "lower-value human capital," not efficiency. Every other bank CFO on that earnings call just got permission to drop the euphemisms in their own transformation decks.

[Sameer Koli @sameergkoli]

The AI bottleneck stopped being model quality, 64% of enterprises now name data and infrastructure as the real constraint, not access to models. In regulated markets, you can't rent your way past that: data residency and governance are architecture decisions, not vendor selection

[Sameer Koli @sameergkoli]

I just published AI Didn’t Kill Code Review. Engineering Leaders Did. https://t.co/14e2SUsuS2

[Sameer Koli @sameergkoli]

RBI's draft AI framework doesn't stop at kill switches. It holds banks responsible for explaining decisions made by vendor models they can't see inside. Most bank-vendor AI contracts have no explainability clause today. The compliance gap isn't the model, it's the procurement.

[Sameer Koli @sameergkoli]

Every enterprise reviews its model vendor's safety card. Almost none review the proxy layer routing every request to it. That's exactly where LiteLLM's chained CVE landed — turning a free open-source dependency into the one place that leaks every provider key you hold.

[Sameer Koli @sameergkoli]

LiteLLM's MCP test endpoints let any authenticated key-holder shell into the host (CVSS 8.7). Chain it with Starlette's Host-header bypass and auth disappears entirely: CVSS 10.0, every provider key behind the gateway exposed. Nobody threat-models the plumbing, only the model.