AI infra as a category tends to get collapsed into three things: foundation models, GPU clouds, and Nvidia. At that layer of the stack (the largest models, hyperscaler compute, the biggest inference platforms), capital and scale do decide the outcome. The players are largely set. It is a hard place for a small team to find a wedge, and that is a fair reading.
Most AI builders, though, spend their day one layer down. In the collection of developer tools, orchestration primitives, observability platforms, evals, and vertical infra that lets them actually build with foundation models rather than just call them. That layer is where the durable software businesses are being built, and it is not concentrated. Teams that ship well-designed developer tools with tight product loops, sharp positioning, and fast iteration are winning meaningful market share. Braintrust reached an 800 million dollar valuation on 120 million dollars raised. Dash0 crossed a billion dollar valuation with 110 million raised, defining AI-native observability as a category. Langfuse was acquired by ClickHouse in January 2026 for what most funds would call a strong outcome. Temporal, a durable-execution primitive, hit a 5 billion dollar valuation on a Series D that closed in February 2026.
None of these companies built a foundation model. Every one of them built a piece of the plumbing that people who build with foundation models cannot function without.
That is the map. India in 2026 has a specific opening in it. The IndiaAI Mission is running 34,000 subsidised GPUs at 150 rupees per hour for startups. Sarvam open-sourced 30B and 105B sovereign models in February. Krutrim crossed 300 crore in FY26 revenue and hit its first annual net profit. Q1 2026 alone brought 3.94 billion dollars of AI funding into Indian startups, roughly six times all of 2025, minting two new AI unicorns in four months. And underneath all of this, a generation of Indian infra engineers who spent the last decade building the payments, streaming, and observability infrastructure of the world is now free to build for AI natively.
The ideas below are the layer of AI infra we think is most buildable in 2026, from India, in a way that a small disciplined team can win. Most of them travel globally. A handful are India-specific because a regulatory or linguistic wedge creates a moat that global players cannot easily copy. All of them share one property: the winners will be developer-love companies, not compute-scale companies.
If you are building one of the twenty below, or a sharper version of one, come talk to us.
1. The India inference cloud
Every foundation model provider (OpenAI, Anthropic, Google, Meta) hosts inference in United States or European data centres. Indian enterprises, particularly in BFSI, government, healthcare, and telecommunications, are under increasing pressure from DPDPA, RBI, IRDAI, and internal risk teams to keep sensitive workloads onshore. The Indian government’s own AI adoption is contingent on sovereign hosting. Neysa is proving early enterprise demand. Yotta, ESDS, and Krutrim are building capacity. But the software layer, an inference platform that behaves like an API to a developer while running physically inside India, is still under-built.
Build an inference cloud designed for the Indian enterprise buyer. Support open-source models (Llama, Sarvam, Krutrim, Mistral) plus a curated set of proprietary alternatives via passthrough. Physical hosting inside Indian data centres with clear DPDPA-compliant data flows. Pricing that undercuts global providers by 30 to 40 percent on the models where hosting matters. SDK that is a drop-in replacement for the OpenAI SDK so developers can switch with a single line change.
Why now: DPDPA rules mature through 2026. RBI’s data residency guidance for BFSI tightened in 2025. IndiaAI Mission subsidises the underlying GPU cost.
Who wins: a founder with real cloud infra credibility (ex-AWS, ex-Google Cloud, ex-Nvidia) paired with an enterprise sales operator who has closed regulated buyers in India.
Watch-outs: this is a capital-intensive business at scale. The seed and Series A can be done leanly, but Series B onwards requires access to GPU financing that few Indian funds are set up for. Have that conversation with your syndicate early.
2. The Indic application layer
Sarvam, Krutrim, AI4Bharat, and a handful of others have delivered foundation model capability in Indian languages. What is missing is the application-ready layer above them: a set of building blocks that any developer can use to add fluent Indian-language capability to a product without training or fine-tuning a model themselves.
Build the Indic application layer. Speech-to-text and text-to-speech APIs across the 22 scheduled Indian languages, priced per minute, with SDKs and clear latency guarantees. Multilingual embeddings tuned for retrieval across code-switched Hindi-English or Tamil-English content. Translation and transliteration primitives that handle Indic idiom, not literal word-for-word. Prebuilt agent skills for the most common Indian workflows (KYC, appointment booking, GST invoice generation, insurance policy explanation).
Why now: Sarvam Bulbul V3 crossed acceptable voice quality in early 2026. Sarvam Audio’s ASR now covers 22 Indian languages. Developer demand for a “just work” Indic layer has crossed a threshold where a well-designed platform can consolidate.
Who wins: a founder with speech or NLP research depth paired with a strong developer-product founder. This is a product-led-growth business, so the developer experience is the entire moat.
Watch-outs: Sarvam and Krutrim are natural competitors here and have the underlying model advantage. Your wedge is the developer surface, the documentation, the reliability, the pricing predictability, and the specific application skills. Compete on those, not on model claims.
3. The production agent runtime
Every fund’s inbox in 2026 has a deck for an agent framework. Most of the current tools are optimised for prototyping and demos, which is a different job from running the same workflow reliably in production. The layer developers need in production, a durable execution runtime that handles retries, state, human-in-the-loop, tool calls, and long-running workflows without falling over, is a category the current leaders (LangChain, LlamaIndex, CrewAI) are still working through. Temporal is doing 5 billion dollar business with durable execution, but Temporal was not designed for LLM workflows.
Build an agent runtime that is native to LLM workflows and durable to production reality. Handles a 40 minute agent conversation that pauses for human input, resumes after a system restart, retries a failed tool call, and produces an auditable trace. Written in a language and paradigm that Indian and global infra engineers already know (Python, TypeScript, Go). Priced predictably, not as a mystery-box cloud service.
Why now: enterprise deployment of agents crossed the pilot-to-production threshold in early 2026. Agent teams that started with earlier-generation prototyping frameworks are actively looking for tools designed around production reality.
Who wins: a founder pair from a distributed systems background (ex-Uber, ex-Netflix, ex-Google infra) plus a developer relations leader with real audience in the AI dev community.
Watch-outs: this category will consolidate. The teams that ship best-in-class in 12 months and land 20 to 50 serious production deployments will win. Slower teams will end up as features inside larger platforms.
4. AI observability and traces
Datadog, New Relic, and Grafana solved observability for microservices. They have not solved observability for AI systems. An agent that calls a model that calls a tool that calls another model that calls a database and returns a hallucinated answer needs a different kind of trace than a REST call graph. Dash0’s billion-dollar valuation shows the market is real; the category is not solved.
Build an AI-native observability platform. Trace the full agent flow from user prompt to final response, including every intermediate LLM call, tool invocation, retrieval, and evaluation. Cost attribution per user, per feature, per agent. Latency debugging that understands where the time actually went. Drift detection for model outputs. Alerting for hallucinations and regressions caught by evals.
Why now: agent deployments are now large enough to justify a dedicated observability spend. Datadog and Grafana are moving into this space but are architecturally constrained by their existing metric-first designs.
Who wins: a founder from a serious observability company (ex-Datadog, ex-Splunk, ex-Grafana) paired with a strong AI systems engineer.
Watch-outs: build the pricing carefully. Metric-volume-based pricing punishes users at scale. Session-based or ingest-flat pricing is the design that keeps large customers happy.
5. Evals as a service
Braintrust’s 800 million dollar valuation on 120 million raised is the clearest signal in the category. Evals are the discipline that separates production AI from demoware, and every AI team eventually gets religion about them. The tooling is still nascent. There is space for a second and third entrant with sharper positioning, better developer experience, or India-specific pricing.
Build an evals platform. A CI-compatible workflow that runs regression, prompt experiments, and canary evals on every deploy. Judge models that are cheaper than GPT-4 class comparisons. Dataset management with human review workflows. Task-specific eval suites for common patterns (RAG, agents, extraction, summarisation, code generation). Integration into the major agent runtimes and CI systems.
Why now: enterprises are moving beyond “the model looks good” to “we can prove regression across 500 evals before every release.” The teams that formalise this discipline are pulling ahead of teams that don’t.
Who wins: an ML researcher or MLE with production experience, paired with a strong developer-tools product person. Developer experience is the entire moat.
Watch-outs: Braintrust, Langsmith, and Weights and Biases will be entrenched. Your wedge is a specific customer segment underserved by them (India enterprise, specific verticals, or specific eval methodologies) plus better pricing.
6. Vector database for the enterprise
Pinecone, Weaviate, LanceDB, Qdrant, and Chroma have battled for the vector database category over the last three years. Most of the winning has happened at the cloud-hosted, developer-first end of the market. The enterprise end, where the customer needs on-premise or hybrid deployment, DPDPA-compliant hosting, and integration with existing data platforms, is under-served.
Build a vector database designed for the Indian and global enterprise. On-premise or hybrid deployment as a first-class option. High-availability, replication, and backup that a bank or a large enterprise CIO recognises. Integration with existing data warehouses (Snowflake, Databricks, BigQuery). Query optimisation for the specific workloads Indian enterprises are running (multilingual retrieval, compliance-heavy search, RAG over policy documents).
Why now: RAG is now a routine deployment inside enterprises. The cloud-only vector DB is a non-starter for a growing set of buyers. The gap is not filled.
Who wins: a founder pair with one deep database systems engineer (someone who has built a real database) and one enterprise sales leader who can move a big-bank procurement cycle.
Watch-outs: Postgres with pgvector is genuinely good and free. Your product has to be 10 times better on the specific workload the enterprise cares about, or you lose to “just use Postgres.”
7. The prompt engineering platform
Every AI team of more than five people ends up building a home-grown prompt management system: a Notion doc, a Google Sheet, a git repo, sometimes a small internal tool. All of them are inadequate. Version control for prompts, A/B testing across models, team collaboration, and safe deployment are not solved anywhere in a production-grade way.
Build a prompt engineering platform. Git-like versioning of prompts, treat prompts as first-class artifacts. Deploy prompts independent of code deploys. A/B test prompts across model providers with statistical rigour. Team collaboration with role-based access. Integration with evals so a prompt change triggers the regression suite. Priced per seat, not per token.
Why now: every AI application team has felt the pain. The winning tool will consolidate the market fast because switching costs are low and network effects around prompt libraries are real.
Who wins: a founder from a developer platforms background (ex-GitHub, ex-Vercel, ex-Notion) paired with a strong AI application engineer.
Watch-outs: this is a feature many observability and eval platforms will bolt on. Your durability comes from being the developer’s daily tool, not the platform team’s dashboard.
8. Data labelling and RLHF as a service
Scale AI, Surge, and Labelbox have proven the market. India has been the physical labelling supply for a decade already, quietly. The opportunity in 2026 is to build the software layer on top of that supply, capturing more of the value chain, and to serve the emerging RLHF and RLAIF workloads that require both scale and quality of judgement.
Build a data labelling and RLHF platform. A trained workforce inside India, but a software layer that serves global AI companies with the workflows, quality assurance, and delivery guarantees they need. Specialised skills (medical annotation, legal, financial, code, multilingual) at global-competitive pricing. Integration into the customer’s training pipeline. Reputation-based reviewer pools rather than gig-worker roulette.
Why now: RLHF and RLAIF have replaced supervised fine-tuning as the dominant training paradigm. The demand for high-quality judgement labels is at an all-time high. Scale AI’s Meta deal in 2024 reset the category’s pricing, opening space for a well-run competitor.
Who wins: an operator with real labelling operations experience plus a strong software founder. This is a business that has to run both a workforce and a platform well.
Watch-outs: quality is the entire business. One bad delivery to a serious customer damages the brand for years. Build the quality process before the sales process.
9. Synthetic data generation
Training data is the most expensive input to an AI company. Real data has legal, privacy, and volume constraints. Synthetic data (LLM-generated or simulation-based) is now credible enough for many use cases, particularly in verticals where real data is scarce or protected (healthcare, financial, autonomous systems). The market has a handful of players (Gretel, Mostly AI, Tonic) and space for more.
Build a synthetic data platform. Generate structured tabular data that matches production statistical properties. Generate document, image, and multimodal synthetic samples for training. Handle the privacy proofs (differential privacy, k-anonymity) that regulated buyers require. Serve specific verticals (financial services, healthcare, autonomous driving) with pre-built pipelines.
Why now: LLM quality is finally good enough to generate synthetic training data that improves rather than harms downstream model performance. Enterprise buyers are actively asking for it as a compliance-friendly alternative to production data.
Who wins: an ML researcher with real training pipeline experience, plus a founder who understands the specific compliance and privacy requirements of the target vertical.
Watch-outs: generic synthetic data is a race to the bottom. Vertical-specific synthetic data with real quality guarantees is a business.
10. DPDPA and PII compliance for AI pipelines
The Digital Personal Data Protection Act came into force in 2024 and its penalties regime is coming into effect in 2026. Every AI application in India that touches user data will need clear consent flows, purpose limitation, right-to-erasure, and audit trails. Global AI application companies also need PII redaction and privacy compliance layers to serve regulated customers. The category exists in the US (Skyflow, Privado, Piiano) and is under-built in India.
Build a data privacy and compliance layer for AI. PII detection and redaction that works in Indian languages (a real gap in existing tools). Purpose-limited data pipelines. Consent management with ABDM-style granularity. Model input and output filtering for confidential information. Auditable data flows for DPDPA compliance officers. Sold to CTOs and CISOs at Indian enterprises and to compliance officers at global AI companies serving Indian users.
Why now: DPDPA penalties activate in 2026. Every serious Indian enterprise CIO has this on their 2026 roadmap.
Who wins: a founder with real privacy or security background plus a strong enterprise sales operator. Compliance sales are relationship-heavy and slow; plan for it.
Watch-outs: do not sell fear. Compliance products that lead with fear-of-fines earn short pilots and lose renewals. Sell operational efficiency alongside compliance and the customer stays.
11. AI security and prompt injection defense
Prompt injection, model jailbreaks, and data exfiltration through AI systems are now real production security concerns. The category has emerging players (Prompt Armor, Lakera, Robust Intelligence, HiddenLayer) but is a long way from mature. Enterprise buyers know they need something in this space and do not know what to buy.
Build an AI security platform. Runtime prompt injection detection. Output filtering for confidential data leakage. Red-teaming as a service, both automated and human-in-the-loop. Model behaviour monitoring for drift and jailbreak attempts. Integration into the enterprise’s existing security stack (SIEM, SOC tools). Sold to CISOs.
Why now: agent-driven autonomy is expanding the attack surface. High-profile prompt injection incidents in 2025 raised board-level awareness. CISOs are budgeting for AI security in 2026 CIO plans.
Who wins: an application security veteran (ex-CrowdStrike, ex-Cloudflare, ex-Wiz) plus a strong AI systems engineer.
Watch-outs: this is a category where the buyer is often skeptical of vendor claims. Build public benchmarks and be willing to be tested. Overpromising once is a career-ending mistake with CISOs.
12. AI FinOps
Every AI-first company past a Series A has a spreadsheet somewhere that tries to track token costs across OpenAI, Anthropic, Google, and open-source hosted alternatives. That spreadsheet is always out of date and always missing at least one line item. FinOps for tokens is a category the market is quietly asking for.
Build an AI FinOps platform. Track cost per user, per feature, per model provider, per prompt version. Alert on cost anomalies. Recommend model swaps that maintain quality while reducing cost. Forecast spend for finance teams. Integrate into the model gateway or observability layer that the customer is already using.
Why now: AI spend has crossed the threshold of being a real budget line item for serious AI companies. CFOs are asking for it. Nobody has built the category-defining product yet.
Who wins: a founder from cloud FinOps (ex-Vantage, ex-CloudZero, ex-Apptio) paired with an AI infra engineer.
Watch-outs: the cost data lives inside the model providers’ APIs, which are inconsistent. Building the ingestion layer well is unglamorous but is 60 percent of the product.
13. AI code review and security
Cursor, Cody, Copilot, and Devin are producing enormous volumes of code. The review layer, particularly for security and quality, is under-invested. Snyk, GitHub Advanced Security, and Semgrep are traditional players. AI code review specifically for AI-generated code is a new category.
Build an AI code review platform focused on AI-generated code. Detect common failure modes (secrets in code, insecure patterns, deprecated APIs, license violations). Suggest fixes with the same AI capability, cutting the cycle time between generation and safe merge. Integrate into GitHub, GitLab, and Bitbucket workflows. Priced per developer per month.
Why now: AI-generated code volume is exploding. Traditional review tools flag too many false positives on AI output. A specialised layer has real demand.
Who wins: an application security founder plus a strong AI engineering leader. This is a very engineering-heavy product; the founder pair needs to have shipped serious platforms before.
Watch-outs: developer trust is fragile. False positives at high volume train developers to ignore alerts. The precision-vs-recall calibration is the entire product.
14. AI DevOps and infra automation
Terraform, Kubernetes, and CI/CD pipelines are the daily reality of every serious engineering team. AI can meaningfully improve the DevOps workflow: infrastructure debugging, incident response, on-call handling, config drift analysis. The category has early entrants (Cleric, Traceroot) and space for a leader.
Build an AI DevOps assistant. Ingest logs, metrics, traces, and infrastructure state. On an incident, generate a hypothesis about the cause, propose remediation, and coordinate the fix with the on-call engineer. On steady state, detect and suggest fixes to config drift, cost bloat, and infrastructure vulnerabilities. Priced per engineer or per infrastructure spend.
Why now: LLMs can now read logs and configs at production scale and reason about them credibly. On-call fatigue is at an all-time high. Engineering leaders are actively looking for tools.
Who wins: a founder with real SRE or infrastructure engineering leadership experience, paired with an AI infra engineer.
Watch-outs: mistakes in DevOps cost customers real money. The product has to be exceptionally careful about autonomous action versus recommendation. Start recommendation-only, earn trust, then graduate.
15. AI test and QA automation
Manual QA is one of the largest labour lines in Indian software services. AI-generated tests, AI-driven regression, and AI-based UI testing are collectively poised to reshape how testing happens. Selenium and its descendants have owned the space for two decades. There is space for an AI-native alternative.
Build an AI testing platform. Generate test cases from feature specs. Run regression across every commit with AI-based UI understanding (not brittle CSS selectors). Prioritise test runs based on the code changed. Report failures with useful, human-readable summaries. Integrate into every major CI system.
Why now: the volume of AI-generated code needs matching test coverage. Traditional test frameworks cannot keep pace. Enterprise engineering teams have budget.
Who wins: a founder from a testing or QA background (ex-BrowserStack, ex-Applitools, ex-Testim) plus a strong AI engineering leader.
Watch-outs: BrowserStack is Indian, well-funded, well-run, and will move in this direction. Your wedge is a specific slice (mobile, or specific frameworks, or a specific developer experience) that they will not prioritise.
16. Voice AI infrastructure
Bolna in India, Retell, Vapi, and Deepgram globally are proving that voice AI is finally a production-ready category. The infrastructure underneath (real-time speech, turn-taking, agent orchestration for voice, telephony integration) is a full stack that most application developers do not want to build themselves.
Build a voice AI infrastructure platform. Real-time speech recognition and generation in the target languages. Turn-taking and interruption handling that feels natural. Agent orchestration integrated with the voice layer. Telephony integration (SIP, WhatsApp Voice, IVR replacement). Pricing per minute or per session. Serve the specific customer needs that a global voice API (Vapi, Retell) will not prioritise: Indian languages, Indian telephony providers, DPDPA compliance.
Why now: voice AI has crossed the acceptable-quality bar in most major languages. Enterprise voice deployments are moving out of pilot in 2026. Vapi hit unicorn valuation in early 2026 on this thesis.
Who wins: a founder with real telephony or voice infrastructure background (ex-Twilio, ex-Ozonetel, ex-Exotel) plus a strong AI engineer.
Watch-outs: telephony integration is the deep unglamorous work that decides winners. Founders who skip it and ship “voice AI on the web” end up as demos.
17. Multimodal generation platform
Image and video generation have moved from novelty to production use in marketing, e-commerce, creator content, and enterprise design. The infrastructure layer (model orchestration, pipeline management, brand consistency, safety filtering) is still being built. Runway, Pika, Ideogram, and a handful of others are competing at the model layer. The application infrastructure layer is open.
Build a multimodal generation platform for the enterprise or the creator business. Model routing across image, video, and audio generation models. Brand consistency guarantees (LoRAs, style transfer, watermarking). Safety and compliance filtering. Version control for generated assets. Team collaboration and asset library management. Sold to marketing teams, e-commerce operators, and content studios.
Why now: video generation quality crossed a threshold in 2025 that made enterprise use viable. Marketing budgets are shifting to generated content. The category has real demand and no clear leader for enterprise workflows.
Who wins: a founder from a creative tools or design platform background (ex-Adobe, ex-Canva, ex-Figma) plus a strong generative AI engineer.
Watch-outs: model quality changes every quarter. Building on top of specific models is dangerous. The abstraction layer that lets customers switch models without changing workflow is the durable business.
18. On-device and edge AI
Consumer devices, industrial IoT, automotive, and healthcare hardware are all pushing AI inference to the edge. On-device inference is faster, cheaper, more private, and often more reliable than cloud inference for specific workloads. The infrastructure layer (model compression, quantisation, deployment, monitoring on the edge) is nascent.
Build an on-device AI infrastructure platform. Model compression and quantisation tuned for specific hardware (mobile SoCs, embedded chips, automotive-grade compute). Deployment and update management across a fleet of devices. Monitoring and observability from edge to cloud. SDK support for the target platforms (iOS, Android, Linux embedded, RTOS). Sold to enterprises with real edge deployments.
Why now: consumer hardware (phones, laptops) is shipping with NPUs capable of usable local inference. Enterprise edge use cases (industrial IoT, connected vehicles) are moving to production. Global players (Qualcomm AI Hub, MediaTek) are opening tooling that a smart third party can build on top of.
Who wins: a founder from an embedded systems or mobile platform background paired with a strong ML systems engineer.
Watch-outs: hardware fragmentation is the trap. Support the top three or four hardware platforms deeply rather than every possible target shallowly.
19. The AI model gateway
Every AI application team past a certain scale ends up building a model gateway: a layer that routes requests across OpenAI, Anthropic, Google, and open-source models based on cost, latency, quality, and availability. LiteLLM proved the demand is real. Commercial products (Portkey, Kong AI Gateway, Cloudflare AI Gateway) are competing. The category has clear space for a leader.
Build an AI model gateway. Route requests intelligently across model providers. Handle failover, rate limits, and cost optimisation. Provide unified observability, cost tracking, and access control. Cache aggressively. Integrate authentication and PII redaction inline. Priced per request or per team.
Why now: AI application teams past Series A have this problem uniformly. The market is aware. The winner will be the tool with the best developer experience, not the most features.
Who wins: a founder from an API platform background (ex-Kong, ex-Cloudflare, ex-Auth0) plus a strong AI infra engineer.
Watch-outs: this is a category where an open-source project could commoditise the paid layer overnight. Your durable business is a hosted offering plus enterprise controls, not the core routing.
20. AI-driven code and system migration
Legacy code migration is a category most people ignore because it is not exciting. It is also a 50 billion dollar global services market that AI is finally credibly attacking. COBOL to Java, Angular to React, SAP customisations to modern platforms, mainframe modernisation, database re-platforming. The category has emerging players (Cognition’s Devin is dabbling, Cognify.ai, GitLab is investing) and enormous latent demand from Fortune 500 IT teams and Indian services companies.
Build an AI code and system migration platform. Ingest a legacy codebase, model its behaviour, generate a modernisation plan, execute the migration in reviewable chunks with human sign-off. Focus on one migration path at a time (COBOL to Java is the classic; Angular to React is the modern equivalent). Sell to Fortune 500 IT teams and to Indian services companies as an accelerant to their existing engagements.
Why now: LLM code understanding at scale crossed the threshold in 2025. Fortune 500 IT teams have modernisation budgets that have been stuck for a decade. Indian services companies are actively looking for AI to protect margins as their base offering commoditises.
Who wins: a founder with real enterprise services background (ex-TCS, ex-Infosys, ex-Accenture consulting arm) paired with a strong compiler or systems engineer.
Watch-outs: this category rewards enterprise sales patience. Nine-month cycles are the norm. The upside is that once you land a Fortune 500 migration, the expansion inside that customer can be enormous.
Picking one
Five filters for AI infra specifically.
First, developer love is the moat. AI infra is a developer-first market. The category winners have been the tools developers pick up on a Friday afternoon, try in fifteen minutes, and champion inside their teams the following Monday. Documentation, SDK quality, examples, and community are not marketing costs. They are the product.
Second, the frontier moves. Every model, every API, every capability that anchors your product will change in six months. Build abstractions and interfaces that survive the change. Founders who bet their product on a specific model, a specific tokenisation scheme, or a specific pricing tier are building a business on top of a fault line.
Third, compete on unit economics. Open source and commodity pressure are both real in AI infra. The categories where the winner is not competing on the marginal cost of a request are the categories worth building in. Categories where you cannot beat GPU-hour math are the categories to avoid.
Fourth, distribute where developers are. Not on LinkedIn. On GitHub, on Hacker News, on the AI Discord servers, at the Bengaluru meetups. Sales starts with product-led adoption. Enterprise contracts follow developer champions. Founders who invert this order end up with a slower sales cycle and a weaker product.
Fifth, India-specific wedges are underrated. DPDPA, RBI data residency, IRDAI on health data, and enterprise appetite for sovereign AI all create real product wedges that a global player cannot easily match. Founders who build for India-first infra needs, then export the same product globally on the same technical base, are compounding two flywheels at once.
A final macro note. The temptation in AI infra is to look at the giants (Nvidia, OpenAI, Anthropic, Databricks, Snowflake) and conclude the category is set. The more useful view is to look one layer down. The businesses that will define the next decade of infra are being built by two-founder teams in Bengaluru, Berkeley, Delhi, and London with fewer than fifty engineers, shipping tools that developers use every day. India in 2026 has the compute, the talent, the regulatory wedges, and the capital to compete in that layer at the level of any global market. What is missing is not the ingredients. It is founders who see the opportunity clearly and start building against it.
This is the last volume in the AI arc. We will follow up with what to build in adjacent categories that have their own compounding stories: manufacturing and B2B commerce, climate and energy, and new consumer brands. If you are building one of the twenty above, or a sharper version of one, we want to hear from you.
A note on intent: this is a thought piece, not an investment thesis.
