Founder @Chingari_in @videoraiq , Projects - @PowerBrowser, @adsgpt_ai, @calorietrackerr @globussoft_ai - Building Enterprise Agentic AI solutions
Bengaluru, Karnataka
Joined May 2015
- Tweets6.9K
- Following5.2K
- Followers13.4K
- Likes10.5K
Sumit Ghosh retweeted
I've been critical in the past of AI compute indexes and the willingness to price compute based on the spot rental price, but now that the AI reset is sinking into the repricing of AI buildout debt its time to realize that the cost of capital is what determines the price of compute.
The most important number in GPU pricing right now isn't $/GPU-hour. It's the gap between what a hyperscaler pays to borrow (~90bp over Treasuries) and what a neocloud pays (~1,100bp).
The cost of compute is now set by the cost of capital, and the spot rental rate no longer tells you much.
The price of an H100 isn't one number. The same chip in the same rack costs a different amount depending on whose balance sheet it sits on. A ~1,000bp funding gap is a structural cost advantage that no amount of operating skill closes.
For example, take one GPU-hour apart. Most of it isn't power or space. It's paying back the chip, the interest on the loan that bought it, and the return the equity expects. A $40–55k all-in GPU, 75% debt-funded, at ~7,000 utilized hours a year is a financing instrument that happens to do matrix multiplication.
As the +500bp on neocloud debt (10%→15%) adds $0.20–0.30 to the break-even cost of every GPU-hour, which is then added on to the rental price.
That's 8–10% of the H100 index, from a credit spread alone, driven almost entirely by the cost of capital.
That means that the compute spot index is a derivative of the AI infra debt credit curve and the squeeze on neocloud margins.
Damn it! Can't wait to play with the open weights man! Congrats @bertgodel !! Lets go kick some ass!
Microsoft shilling Chinese open weight models !! What day in the history of computing 🤣🤣 once known for closed source shitty OS now shilling open weight models to sell their PCs 😀😀
Microsoft today showcased DeepSeek V4 Flash, a 284B-parameter open-weight model, runs locally on a Windows PC once quantized to 1.6 bits, which brings it down to about 60GB of memory.
Their new Surface Laptop Ultra, built on Nvidia's RTX Spark chip with up to 128GB of unified memory, has room for it in its higher-memory configurations.
----
Full video on "Windows" YouTube channel, (link in comment)
I built a content studio where AI agents do the work, and I am the only human in it.
Its called Optra Studio. Here is what it does today:
🤖 Agents that work with you, 24/7
A main agent runs everything. Other agents use the dashboard the way a person would. They plan, check and fix, and when something breaks they report it so another agent can fix it.
✍️ From story to storyboard, Claude writes the story. A local Qwen model plans the scenes, adds the narration and does the storyboarding, so that part never leaves our own GPUs.
🎨 Pictures, voices, motion, cut
• Drawing: Flux
• Voices: Kokoro, Chatterbox, and Audar for Arabic
• Animation: Wan and LTX
• Cutting: our own FFmpeg-based editor
🌍 Two languages - Every episode comes out in English and in Arabic.
✅ Checked before you see it. Every shot is reviewed, and every frame of every clip is checked at 24 fps.🎥 Your own footage too. Upload your videos and edit them with GenAI.
💸 No credits to count
Unlimited generations for a fixed monthly fee, plus a limited number of free generations to start.
Already made in it: Stories of Wulfa, an animated series in English and Arabic, and Flat Nine, a short.
Try it at studio.optracloud.com: create a project and upload your videos.
#AI #GenerativeAI #AIAgents #ContentCreation #Animation #VideoEditing
yup and it will only get better from here! Business who cant talk to agents will get left behind!
Personal AI Assistants like Instinct and Meta Muse will help users avoid the struggle of navigating through poorly developed User Interface of Websites.
Today I ordered something from Decathlon without having to struggle through their website. I don't like to share my Credit Card details with Instinct and so, it shared the UPI QR screenshot with me and I made the payment easily.
Crazy times!
Sumit Ghosh retweeted
Extremely bullish on $5,000/month managed AI agents as a service.
The agent:
-one defined job (speed to lead, quote prep, invoice chasing)
-its own hosted setup, so the client never touches anything technical
-tools connected once through Composio, one API key instead of a folder of logins
-skills written from client call transcripts/a structured assessment with the client
The service:
-hosting it
-fixing it before the client notices it broke
-adding capabilities every time they ask
-a human who answers when something looks wrong
The math
-$5,000/month is about half the cost of a quality domestic employee
-it doesn't take leave, doesn't complain, never clocks out
@nickvasilescu runs this model. The key here is that the client is paying you to deliver a consistent, repeatable, high-quality OUTCOME. They don't care that an agent is the one delivering it (or anything about the tech at all for that matter).
Sumit Ghosh retweeted
How to get $1M in rev from Gemini recommending your start-up. Companies pay me $500k / year to advise on this.
1/ Find the topics you want to rank for (e.g. "PDF -> Word doc converter")
2/ Turn that topic into 3 query intents. Such as:
- "What is the best PDF -> Word doc converter"
- "How to convert a PDF to a word doc"
- "[YOUR BRAND NAME] review for PDF -> word doc conversion"
3/ Record a 90 second Youtube video for each one. If you're lazy, use AI to make it. We find that @trymelius is quite good for this stuff.
4/ Post the videos. Make sure you have a good transcript. Repeat.
Youtube is the most under-utilized AEO tactic today.
Who this works especially well for:
[] Niche industries
[] B2B SaaS
[] Smaller brands with limited footprint today
Sumit Ghosh retweeted
Qwen Image 2.1 is now open weight at just 7B parameters!!
It beats Nano Banana 2, Pro and Grok Image. This is finally the image model update I’ve been waiting for, with weights anyone can actually run. So excited for this, thank you @QwenDevs!
Yes!!!
UPDATE: Kev-0.6B, 4B, and 8B are now available. Kev is a family of small open source Jev-like decision models you can train and run yourself.
This new family is based on Qwen3 using the same LoRA + small pointer head technique as before, but scaled up.
Out of domain, on data Kev never trained on: Kev-8B 79.6%, Jev 85.7%.
• Drop-in TypeSafe System One API; their SDK works with one `base_url` change
• Kev-4B serves on a 32 GB Mac in bf16: ~300 ms for five questions, ~40 ms on an H100
• Repeated documents hit a KV cache: 2-2.5x faster
• Apache 2.0 License. Kev-4B trains in 40 minutes on one H100. Kev-8B in 83 minutes.
Code, weights, evals: github.com/jaredpalmer/kev
Sumit Ghosh retweeted
Monetization of software is dead now.
Anyone can clone your app in few hours, copy your ads, ditto your positioning.
I see a new opportunity now.
Monetization of SLMs (small language models)
You can fine-tune 0.5B to 12B models for specific use cases.
Then sell licenses to these models like you do for SaaS.
Models like Deepseek v4 pro, GLM 5.3, Kimi k3 or even GPT 6 Astra are really smart at understanding fine tuning jobs.
From preparing the dataset, LoRA adapter, to exporting the model as GGUF or MLX, these models can do end-to-end jobs for you.
I've fine-tuned 20+ models in last pne year and I'm surprised that not many people on X talk about this.
A fine-tuned SLM can beat a SoTA model on specific use cases.
That can drop API costs by 95%, make it 5x faster, and its your model, hosted privately.
Custom SLM > SaaS now. Dive into this space.
Congrats @julien_c !!
Super happy to officially announce that we are a̶c̶q̶u̶i̶r̶i̶n̶g̶ joining forces with @nvidia 🔥
Here is a more personal take:
AI is at an inflection point. Open source AI can become less relevant in the coming years if the big closed labs run away with it, OR it can become the foundational fabric of the next phase of human civilization.
Those are vastly different outcomes, and we need the critical mass to ensure we give our collective best shot to the second outcome. Given @JensenHuang's stance on open source AI and how he stepped up to defend it when it was under threat earlier in the summer, NVIDIA was the only partner we truly considered.
HF will remain an independently run, neutral platform.
This gives fuel to our long-term vision and mission of unlocking the community's progress to ensure that AI, which is the greatest breakthrough of our lifetime, is accessible to as many people as possible.
Sumit Ghosh retweeted
Received several pings today from the ecosystem to understand this better and what it means for Indian semiconductor startups. Here is a simplistic thread on the AI hardware stack, @C2iSemi and India's role within it.
6 months from Series A to M&A - I didn't expect to use this picture this soon :)
Congratulations to Ram, Preetam, Vikram, Dattatreya, and the entire @C2iSemi team! It was wonderful partnering with you folks.
moneycontrol.com/technology/…
ALT At the C2i office post partnership
Sumit Ghosh retweeted
NVIDIA has lost it 🤯
Someone just made CUDA run on Apple Silicon.
Your MacBook can now run the same particle simulations NVIDIA H100 clusters do, using the EXACT CUDA code they wrote.
For years, CUDA only speaks NVIDIA.
Apple GPUs speak a different language called Metal. So they built a translator chain using get 5.6 sol that rewrites CUDA 5 Metal step by step through existing compilers.
zero source code changes.
They tested it on a heavy 3D fluid simulation, not a hello-world demo. this one sorts particles, builds neighbor lists, and runs atomics simultaneously. the kind of workload that breaks toy translators.
Results on a single laptop chip:
→ Apple GPU: 6 seconds
→ Same code on CPU: 60 seconds
→ 10x faster, GPU fully saturated the entire run
The translation layer wasted zero performance. cross-compilation usually eats 30–50% of your throughput. this eats none.
Every CUDA codebase in the world just became runnable on a MacBook. the NVIDIA moat everyone said was uncrackable just got a crack.
100% open source.
No argument to this sir!
Stripe bought OpenRouter for $7.5B this week and everyone just shrugged. That's 7x what Instagram sold for. The OG crew is back on More or Less and we got into it:
— My take: there is NO technology in OpenRouter. It's a key swap with great marketing. Ramp announced a router. Everyone will have a router. Stripe paid $7.5B for volume and a look at everyone's request pile — which, to be fair, is worth something.
— OpenAI is the greatest fundraising story in the history of the world. But the 'one winner in AI' narrative is dead. Leads don't matter, switching is free, and nobody can price these companies. If you told me I had to hold Anthropic stock for 5 years, I genuinely don't know what I'd pay. Neither does anyone else.
— Dave's take: the AI backlash isn't about data centers or water. It's that regular people think AI sucks. Slop in their inbox. Nobody in Montana cares about AGI timelines.
— And when the IPOs come and number goes down? Pensioners hold the bag while 50 people walk around SF with $40B each. Second and third order effects are going to be wild.
Come argue with us.
8 million advertisers now use AI to make their ads.
That number doubled in a year.
Which means the capability everyone spent last year racing to build is now a button anyone can press. Generating creative isn't a moat. So what is?
Here's the uncomfortable math — volume without a feedback loop is just faster waste:
→ On Meta, most concepts follow a ~10-day decay curve. Peak on days 1-3. Fatigue by day 5. A 30-50% CTR drop by day 10.
→ Meta's Andromeda ranking now burns through a concept in 2-3 weeks that used to last six.
→ Roughly half of all creatives are retired before day 28 — measured across $1.3B in ad spend.
Now put a human on top of that. Most teams check campaigns once a day. Plenty check on Monday.
Your ads decay hourly. You review them weekly. Everything in that gap is burned budget — and it's a big reason ~30% of digital ad spend still produces nothing.
So the moat isn't making 100 ads. Anyone can make 100 ads now.
The moat is the loop: notice what's dying, kill it, move the money, learn, repeat — faster than the decay curve. That loop is the only thing that compounds. It's also the part nobody wants to do at 2am on a Sunday.
That's why we built Autopilot into AdsGPT.
It isn't "AI writes your copy." It's the layer that runs the loop on your Meta ads while you sleep:
→ Hourly audits against 37 performance rules. Not daily. Hourly.
→ AI Audit calls what's fatiguing, what's working, what to cut.
→ Then it acts: pauses losers, pushes budget toward winners.
But here's the part I'd care about most as an advertiser: automation you can't trust never gets switched on.
So every action passes safety gates. Dry-run mode lets you watch it work before it touches your budget. Every decision is logged. And anything it does, you can undo.
Autonomy without the blindfold.
The last decade of advertising rewarded whoever had the biggest budget. The next one rewards whoever closes the loop fastest — because by now, everyone already has the creative.
Try it free — 35 creatives, no credit card: adsgpt.io
Honest question for the advertisers here: how often does someone actually look at your campaigns? Daily? Weekly? Be honest.
#AI #AdTech #MetaAds #PerformanceMarketing #MarketingAutomation #GrowthMarketing #Startups
In three years, the highest-performing marketing "team" at most companies won't be a team.
It'll be two people and a fleet of AI agents.
Look at what the data already says — not in 2030, right now:
→ Global ad spend grew 8.6% last year. Agency holding-company revenue fell 1.2%. The money didn't shrink. It just stopped flowing to big teams.
→ 32% of brands plan to bring ALL creative in-house within 12 months. 60% of senior marketers have already cut agency spend because of AI.
→ The classic retainer is quietly bleeding out — 20-30% of its value gone, no cancellation email required.
For 30 years, marketing scaled by adding headcount. A designer. A copywriter. An editor. A media buyer. A dozen people to carry one idea across six platforms.
That model is breaking — not because marketers got worse. Because the economics flipped.
87% of marketers now use generative AI, up from 51% in 2024. The average marketer saves 6+ hours a week. Teams report ~44% higher output. When one person with the right agents does what five used to, the org chart stops being an asset and starts being overhead.
Here's the part most people miss: this isn't "AI replaces marketers." It's "small teams with agents beat big teams without them."
The job changes shape. You stop producing and start directing. Taste, brand and judgment — the things AI can't fake — become the entire job.
That fleet isn't a someday idea. It's what we built with AdsGPT.
Not a copywriting toy. Not one tool cosplaying as an agent. AdsGPT is a coordinated fleet of specialized agents that work together like a real growth team:
→ A creation agent studies 500M+ live ads to find what's winning, then turns one brief into 100+ on-brand image and video ads.
→ A scaling agent launches the winners straight to Meta and pushes budget toward whatever's converting.
→ A monitoring agent watches every campaign around the clock — catching fatigue, killing losers, flagging what to scale.
One brief goes in. The agents hand off to each other and carry it from idea to live, self-optimizing campaign — while you sleep.
The proof is in the numbers: brands on AdsGPT ship ~2.8x more creative, hit up to 4.8x ROAS on winners, and cut production costs ~80%. More than 10,000 marketers are building with us — including our own team at Chingari.
What I actually believe: within a few years, every brand becomes a media company. The creative team becomes an AI fleet supervised by a human — not replaced, amplified. Strategy and taste stay human. The grind does not.
Try it free — 35 creatives, no credit card: adsgpt.io
What's the one creative task you'd hand to an AI agent tomorrow?
#AI #Marketing #AdTech #GenerativeAI #FutureOfWork #GrowthMarketing #Startups
"We're hiring a Performance Marketer."
I see 20 of these a day. And more and more, my honest reaction is: you might not need to.
Not because the work stopped mattering — it matters more than ever. Because the work itself is being unbundled.
Look at what a performance marketer actually does day to day:
→ Spin up and test ad creative
→ Run media buying across Meta, Google, TikTok
→ Read the data — kill losers, scale winners
→ Repeat, relentlessly
For 15 years that took a specialist (or three). Today most of it can be run by AI agents — not "tools a human operates," but agents that own the loop end to end.
This isn't theoretical: over a million advertisers already generated 15M+ AI-made ads in a single month on Meta alone. The execution layer is going autonomous.
That's the bet behind AdsGPT: AI agents that do performance marketing — media buying, creative, testing — so a founder or a generalist can drive real paid results without hiring a specialist to run it.
What doesn't change: strategy, positioning, brand taste, knowing which bets to place. That stays human. The agent handles execution and volume; you own the judgment.
So the 2026 question isn't "who do we hire to run performance?" It's "what should a human own, and what should an agent run?"
If you're about to post a "Performance Marketer wanted" ad — pressure-test the role against that question first. You might deploy the budget very differently.
Would you hand your paid media to an AI agent today? Where's your line? 👇
