Manmit Walia

Managing Partner at SourceMash. We build AI enablement, automation & digital transformation that ships - real strategy plus execution. Let’s Talk - SourceMash

[Manmit Walia @manmitwalia]

Two more of Google's top Gemini researchers are reportedly off to Anthropic. The headline is the talent war. The real story for buyers: frontier capability is consolidating into 2-3 labs — so your "AI strategy" is now a vendor-concentration bet. Plan for it.

[Manmit Walia @manmitwalia]

The best-run engagements I've been part of all start the same way: hour one isn't about the work. It's about agreeing how we'll disagree — who decides when we're stuck, how fast we escalate. Settle that on day one and you never lose a week to a standoff in week six.

[Manmit Walia @manmitwalia]

4/ Surface the Q&A. Real customer questions plus your answers are the most citable content on your site — they match how people ask AI. Stop hiding them three clicks deep.

[Manmit Walia @manmitwalia]

1/ Your customer reviews are now SEO assets. AI answer engines pull from them when deciding what to recommend. Most brands treat reviews as a vanity widget. Here's how to make them work for search instead. 🧵

[Manmit Walia @manmitwalia]

3/ Structure them so they're readable. Reviews trapped in a JS widget an engine can't crawl don't exist. Render the text in HTML, with the product and the attribute clearly tied to it.

[Manmit Walia @manmitwalia]

The most expensive consultant is the one who agrees with you in the first meeting. You're not paying for a yes. You're paying for the person willing to tell you the project you scoped isn't the project you actually need.

[Manmit Walia @manmitwalia]

M-Files just shipped AI agents for document workflows. The tip nobody sells you: a content agent is only as good as your metadata. Messy structure in, confident-but-wrong out. Fix how your files are tagged and organized before you point an agent at them.

[Manmit Walia @manmitwalia]

OpenAI is building its own inference chip with Broadcom. The signal for buyers: the cost that matters isn't training the model — it's running it. When the biggest lab optimizes inference economics, your per-seat AI pricing follows. Budget for the run, not the demo.

[Manmit Walia @manmitwalia]

SpaceX buying Cursor for $60B isn’t a coding story.

It’s a stack story. Whoever owns the compute AND the developer’s cursor owns the whole pipeline — from the chip to the keystroke.

The platform war just moved up a layer. “Best model” stopped being the moat a while ago.

[Manmit Walia @manmitwalia]

We stopped asking teams “is it done?”

Now we ask: “what would you need to feel safe shipping this?”

The first question gets you a yes. The second gets you the real risk — the thing they were quietly worried about but no one made space to say.

[Manmit Walia @manmitwalia]

4/ A checkout an agent can actually complete — fewer forced steps, fewer human-only gates on the path. And copy specific enough for a model to quote back. Optimize for the agent’s shortlist, not just the human scroll.

[Manmit Walia @manmitwalia]

1/ Agentic commerce is almost here. Soon a shopping AI — not a shopper — lands on your product page, evaluates it, and decides whether to buy. Most stores aren’t readable to a machine that’s trying to check out. 4 things to fix: 🧵

[Manmit Walia @manmitwalia]

3/ Real-time inventory + price via API. Agents won’t tolerate “add to cart, then discover it’s out of stock.” Stale data gets you dropped from the shortlist on the spot.

[Manmit Walia @manmitwalia]

Most “strategy decks” aren’t strategy.

They’re an expensive permission slip for a decision the client already made.

The real work isn’t the 40 slides. It’s the one conversation where someone finally says the thing everyone was avoiding.

[Manmit Walia @manmitwalia]

Another AI “modernization platform” just raised $60M this week.

These tools are getting good at the rewrite. They still can’t tell you WHICH legacy system to modernize first.

That’s a discovery question, not a tooling one. Modernizing the wrong one fast is just expensive.

[Manmit Walia @manmitwalia]

AI’s real 2026 bottleneck isn’t model quality. It’s compute supply.

When the labs behind your tools are paying over a billion dollars a month just to rent capacity, that cost lands somewhere — usually your renewal.

Budget for AI like a volatile utility, not a fixed SaaS line.

[Manmit Walia @manmitwalia]

OpenAI raises $122B at an $852B valuation — and Nvidia put $30B of it in.

The headline is the number. The real story is the circle: the chipmaker funding its biggest customer to go buy more chips.

That's not investing in AI demand. That's manufacturing it.

[Manmit Walia @manmitwalia]

A thing I underestimated for years: protecting the team's focus is a leadership job, not the team's.

Every "quick favor" you wave through is a tax someone pays later in context-switching.

The best operators I know are ruthless about what they keep OFF the team's plate.