Waqar Uddin

Picking Battles in the AI Stack

May 2, 2026 (5d ago)252 views

A few weeks ago I wrote about how alternative clouds, neoclouds, and telco operators are building credible AI offerings without owning the full stack. The point was that you don't need to play at every layer. DigitalOcean built a real inference business at one layer. Akamai bet on the edge. Hetzner just sells boxes. Each picked their fight.

Then I read Chamath's piece on the AI stack, which makes a sharper version of the same argument from a very different starting point. He maps AI into six layers: infrastructure, chips, data, models, execution, and application. At every layer there is what he calls a fulcrum asset, a chokepoint that confers near-monopoly leverage. ASML at the lithography step. NVIDIA at the GPU layer. A specific Japanese chemicals company for photoresist film. Whoever controls the fulcrum, he argues, captures disproportionate value for the next forty years.

It is a useful frame. It also forces a question I think people in Pakistan keep dancing around. Of those six layers, which ones can a country like ours realistically own? Not aspire to. Not write policy documents about. Actually own.

I went layer by layer. The honest answer is three of the six.

Disclosure: I work at Jazz as Principal Evangelist Cloud & AI. This article is written as personal industry analysis. Views are my own.

Layer 1: Infrastructure

Partial leverage. We have land. We have grid capacity, at least in pockets, though our power tariffs and reliability story is uneven. We don't have rare-earth mineral deposits or a critical materials industry. Our edge here is narrow. We can host data centers. We cannot own the supply chain that fills them.

Verdict: useful but not a fulcrum. Build for it pragmatically, but do not expect to extract monopoly rent from being a power-and-cooling provider. The hyperscalers will outbid us on every megawatt that matters, and the upstream supply for transformers, switchgear, and high-density cooling sits with a handful of European and East Asian firms we cannot displace.

Layer 2: Chips

Zero leverage. There is no scenario in the next decade where Pakistan owns any meaningful piece of the chip layer. ASML, TSMC, Samsung, NVIDIA, AMD, the whole thing is locked up by capital intensity, IP, and geopolitical permission. The fulcrum at this layer was decided twenty years ago and is being defended by export controls.

I bring this up because there is occasional noise in our policy circles about Pakistan building a chip strategy. That noise should stop. The strategic question for us is not how to make chips. It is how to import them at acceptable cost, run them efficiently, and put them to productive use. A 48% import duty on GPU servers does more damage to our AI prospects than a hundred policy white papers can repair.

Verdict: not a battle to fight. Buy, do not build.

Layer 3: Data

This is where it gets interesting, and where I think we have actually been underselling our position.

Chamath frames data as a layer like any other, with its own chokepoints. For a country, the chokepoint is regulatory and linguistic.

Regulatory first. SBP has already drawn a line on where regulated banking data can sit. The pending Personal Data Protection Bill will draw similar lines for personal data. NEPRA is moving in the same direction for energy. None of this data can leave the country in any straightforward way. The hyperscalers either come to us, or they get replaced by something that already lives here. That is a structural fulcrum, defended by law rather than by capital, and it is not going to weaken.

Linguistic second. Urdu, Pashto, Punjabi, Sindhi. Together this is a language footprint of more than 250 million speakers, and not one of them is a priority for OpenAI, Anthropic, or Google. The local language data corpus is a layer no foreign provider can pretend to own. Whoever assembles, cleans, licenses, and continuously refreshes the largest Urdu and regional language corpus is sitting on a fulcrum asset for the entire South Asian AI value chain. The NUST and Jazz partnership on a local LLM is a small step in this direction. There should be ten more like it.

Verdict: this is the strongest fulcrum we can plausibly own. Domestic regulated data plus regional language data is a moat foreign providers cannot legally or commercially cross.

Layer 4: Models

Mostly low leverage at the foundation tier. Pakistan is not going to train a frontier foundation model. The compute economics rule it out. A single training run for a frontier-scale model runs into hundreds of millions of dollars now, and rising. Even at the floor of that range, no country our size is going to allocate that capital toward one training experiment, let alone the dozens of attempts a competitive lab actually needs.

But foundation models are not the whole layer. Fine-tuned models, domain-adapted models, and small open-weight models running on top of someone else's pretraining are very much within reach. A bank-tuned Qwen. A medical-records-tuned Mistral. An Urdu-conversational Llama variant. The model layer above the foundation tier is fragmenting into thousands of vertical specialists, and the cost of running a model has dropped roughly 1500x in six years, as Chamath notes. That curve is in our favor, not against us.

Verdict: ignore the frontier race. Build vertical and language-specific models on top of open-weight foundations. The compounding here happens in fine-tuning data, not in pretraining FLOPs.

Layer 5: Execution

High leverage. This is the layer where physical presence, integration capability, and regulated industry knowledge matter most, and it is the one I think Pakistan most consistently underrates.

Execution is the unglamorous middle of the stack. Taking a model, deploying it on infrastructure, wiring it into a bank's core systems, getting it past compliance, monitoring it in production, retraining it when drift happens. None of this can be done over Slack from Mountain View. It requires people who understand the local regulator, the local enterprise architecture, and the local procurement process. It requires being able to walk into a meeting at the State Bank and explain why the inference logs satisfy a particular section of BPRD Circular 01/2023.

We have those people. Pakistan has more than 300,000 software professionals and a deep system integration culture going back two decades of bank-tech and telco projects. The execution layer is where local advantage compounds and offshore providers structurally lose, and yet it is the layer most under-discussed in our AI policy conversations.

Verdict: this is a fulcrum we already partially own and should consolidate hard. The companies that can deliver production AI deployments inside HBL, OGDC, NADRA, and the federal ministries will own a position no hyperscaler partner team can replicate.

Layer 6: Application

High leverage, with caveats. The application layer is where local context turns into product. An Urdu legal-research assistant. An agricultural advisory app for cotton farmers in Multan. A KYC and fraud package tuned to CNIC patterns. A maternal health chatbot in Sindhi for rural deployment. These are problems no foreign vendor will prioritize.

The caveat is distribution. Building local apps is one thing. Reaching scale is another. This is where telco distribution, mobile money rails, and existing customer bases become the multiplier. The reason I think the telco-AI thesis is structurally stronger here than anywhere else is that the application layer needs a delivery channel, and JazzCash, with its 58 million customers and 20 million monthly active users, is one of the few channels in the country that can reach that scale without inventing a new go-to-market motion.

Verdict: layer six only matters if you can distribute. Pair application development with existing distribution rails or you lose to whoever does.

What this changes

If you take the six-layer view seriously, the strategy more or less writes itself. We stop pretending we are going to build chips or compete on foundation models. We pour resources into the three layers we can actually own. Domestic and language data. Execution capability. Locally distributed applications. Infrastructure we build pragmatically, as plumbing for the other three.

This is not a defeatist read. It is the opposite. Three layers of the stack, controlled at scale by a country with sovereign data, regional language depth, and a working distribution rail, is more strategic position than most countries our size will ever have. The mistake would be to dilute that position by chasing layers we cannot win.

Chamath's frame ends with the claim that fulcrum control compounds. Whoever owns the chokepoint at a layer keeps owning it as the layer matures. That is the part I want our policymakers and CSP CEOs to internalize. The data layer fulcrum for Pakistan is being decided right now, in the next two or three years, by who gets the regulated workloads, who builds the Urdu corpora, and who licenses the language models that come out of them. After that the position calcifies.

We have a window. Not at every layer, but at three of them. That is enough to build a real AI economy. It is not enough to win the layers we keep daydreaming about.

Part of an ongoing series on Pakistan's cloud and digital infrastructure from a practitioner's perspective. Previous posts: Beyond Hyperscalers, Pakistan's 5G and Sovereign Cloud Convergence, and the SBP regulatory framework series.