Krish Subramanian

My post for The Economist reflects on Elon Musk's escape from the bounds of shareholder capitalism through the lens of Hirschman's marvelous history of pre-capitalism, The Passions and the Interests: "Before anyone claimed that markets were efficient, political thinkers claimed that the self-interest of merchants would be a gentler master than the passions of princes.... But we have arrived somewhere Montesquieu and Adam Smith did not foresee. The self-interest of merchants did not tame the passions. It became infused with them, and turned the greatest of merchants back into princes carrying all the undisciplined appetites of old. They built a new and stranger tyranny: not the old corporate machine obedient to shareholders, but a machine that uses shareholder capitalism’s legal forms while escaping its restraints." economist.com/by-invitation/…

Agentic AI does not break because the model is weak. It breaks because the system around it was built for chat. The useful part of Google Cloud's 2026 State of AI Infrastructure report is not the headline that 83% of organizations need upgrades for agentic AI. The useful part is why. Agents change the unit of work. A chatbot is usually one prompt, one model call, one answer. An agent is a loop: plan, retrieve context, call a tool, observe the result, update state, call another tool, maybe retry, maybe ask for approval, then produce an output. That sounds like product magic. Operationally, it is a tax. Google surveyed more than 1,400 senior IT leaders. CIO Dive reported that only 17% have full confidence their stack can support mission-critical agents. Google says 62% are seeing high inference costs from data egress, storage bloat, and idle specialized hardware. It also says 81% cite operational complexity as a hidden scaling cost. This is the mechanism most AI business cases still miss. The cost is not just tokens. It is tokens plus context movement plus retrieval plus orchestration plus retries plus auditability plus latency. That is why "the model got cheaper" does not automatically mean "the workflow got cheaper." If one workflow now triggers hundreds of downstream actions, the price per token can fall while the cost per completed task rises. The CIO question should change from "Which model are we standardizing on?" to "What is our cost and reliability per completed workflow?" That forces a different architecture discussion: 1. Route each task to the right model and silicon. 2. Keep context close to the workflow. 3. Track agent traces, not just app logs. 4. Measure retries and tool failures as cost drivers. 5. Treat approvals and permissions as part of the runtime, not a policy document. The winners will not be the companies with the most agents. They will be the ones that can run agents with predictable cost, latency, reliability, and blast radius. That is less glamorous than a demo. It is also where production happens. #EnterpriseAI #AIAgents #AIInfrastructure #CloudComputing #PlatformEngineering #CIO #MLOps

Apple vs OpenAI is not just a lawsuit. It is a preview of the next platform war. Apple sued OpenAI and two former employees on Friday, alleging trade secret misappropriation tied to OpenAI's consumer hardware ambitions. Reuters reported that Apple claims more than 400 former Apple employees now work for OpenAI. Axios reported allegations that a former Apple engineer accessed Apple's cloud file storage after leaving, and that OpenAI hardware leader Tang Tan allegedly used Apple codenames and asked candidates to bring Apple parts to interviews. OpenAI denies interest in Apple's trade secrets. The legal facts will take time. The strategic signal is already clear. AI companies do not want to live forever inside someone else's device, app store, browser, operating system, or assistant layer. If the interface to the user becomes the agent, then the company controlling the agent gets the customer relationship. That is why hardware suddenly matters again. Not because everyone wants to make a better phone. Because the phone is the current choke point between users, context, identity, payments, notifications, sensors, and daily habits. If OpenAI can build a device or companion layer where agents replace apps, it reduces dependency on Apple. If Apple preserves control of the device layer, it can decide how much of the agent era flows through its rules, permissions, and economics. This is not only about consumer hardware. Enterprise buyers should pay attention because the same pattern will repeat inside companies. Every AI vendor wants to become the system of action. Every incumbent platform wants to remain the control point. The fight will move from "which model is smarter?" to "which layer owns workflow, identity, memory, permissions, and distribution?" The practical lesson for CIOs is simple: do not let one vendor quietly become both the assistant and the operating layer for critical work. Separate the layers where you can: identity, data access, model choice, orchestration, audit logs, and user experience. Vertical integration can create great products. It can also create switching costs before procurement even notices. The Apple-OpenAI case may be about alleged trade secrets in court. In the market, it is about who owns the post-app interface. #AIPlatforms #EnterpriseAI #OpenAI #Apple #AIAgents #TechStrategy #CIO

Vinod Khosla is buying SeaHawks 🤔

Every enterprise leader should balance token costs against governance costs before even comparing the costs with benefits.

Dedicating it to all the Libertarian tech bros who supported Trump. Loving it!! wsj.com/tech/the-white…

The agent is moving from chat window to work surface. OpenAI's ChatGPT Work launch is not really about another model release. The model matters. GPT-5.6 Sol, Terra, and Luna are positioned around better performance per dollar, with OpenAI claiming Sol is 54% more token efficient on agentic coding tasks. That matters because enterprises are finally asking the right question: not "how smart is the model?" but "how much useful work do we get per dollar?" But ChatGPT Work is the more important signal. OpenAI is trying to turn ChatGPT from a place where employees ask questions into a work surface where agents gather context across apps and files, create documents, slides, spreadsheets, and sites, and stay with a project for hours. The examples are not science fiction. Month-end variance analysis. Product launch checks. Sales prep. Conference follow-up. These are the messy middle-office workflows where a lot of enterprise time disappears. Day 2 is where this gets hard. If an agent can touch Slack, SharePoint, Salesforce, Jira, email, browser sessions, and local files, the real product is not just intelligence. It is permissioning, audit trails, spend controls, source grounding, action approval, and rollback. Otherwise, the same thing that makes the agent useful also makes it dangerous. My buyer question would be simple: Can you show me the full chain from source data to final artifact to approved action, including who gave access, what changed, what the agent inferred, and what it was not allowed to do? If the answer is a demo, wait. If the answer is an operating model, pay attention. #EnterpriseAI #AIAgents #AITransformation #FutureOfWork #CIO #PlatformEngineering

Wanting to learn Interpretability of AI is ok but expecting that AI should use the same type of reasoning spaces understood by humans makes little sense. Similarly, the human urge for alignment is another area which makes little sense to me (unless, of course, I put human ego to be at the center). For these reasons, I have a feeling that all the research we are doing to poke into the blackbox of AI is similar to the alien search we have been conducting so far. Maybe, just maybe, we are looking in the wrong direction.

July 13th marks the day when many @AnthropicAI users will move towards @OpenAI for complex cognitive tasks after the removal of Fable from subscription plans

@PeterDiamandis @alexwg I articulated the same point about Meta which Alex said in the recent Moonshots podcast. Meta will come back to regret their decision to move away from OSS cnbc.com/2026/06/14/met…

The crybaby Alex Karp should STFU. He reminds me of legacy companies circa 2010 whining about AWS. If his company is going down, it is because of their strategic decisions. Companies like Cisco, IBM and other legacy companies managed to survive the cloud era (of course, with their brute force marketing). If he is smart, he will use their strategy to stay relevant in the AI era. Whining and FUD is not helping him or his company.

@AyvoLife I 100% agree. That is one of our core focus area with @HikigaiInc platform

@krishnan The trust layer under healthcare software is the hard part. Solid compliance and audit trails make patient decisions clearer. Building that base once beats every team reinventing it.

Let us be clear. Iran MoU was poorly designed with vague wordings in order to announce on Trump's birthday. Now Vance will take the blame. I have no sympathy for either one of them but I am pointing out to the pattern just like any ML algorithm will do

@ashleymayer My obituary reads I used to wait for the ecommerce website to load (read dialup) 😅

Please put it in my obituary that I never waited in a line outside a store.

The @NVIDIAAI and @LangChain NemoClaw announcement is a useful reminder: agents are not won at the model layer alone. LangChain says its new NemoClaw Deep Agents blueprint with NVIDIA combines Nemotron 3 Ultra, LangChain Deep Agents Code, and NVIDIA OpenShell. In its eval suite, LangChain reports an aggregate score of 0.86 at a cost of $4.48, compared with $43.48 for the next closest performing model. Take vendor benchmarks with the usual caution. But the mechanism matters. The agent stack has at least three separable layers: 1. Model: reasoning, instruction following, context handling. 2. Harness: planning, memory, tool use, retries, eval loops. 3. Runtime: sandboxing, policy, observability, deployment control. Most AI debates still over-index on layer one. Production failures usually happen in layers two and three. That is why lower inference cost is not just a CFO story. It changes engineering behavior. Teams can run more evals, specialize agents by workflow, test more failure modes, and keep more traces. Cheap tokens can become better governance if the system is designed that way. My read: the next serious enterprise AI platform will not be the one with the fanciest demo. It will be the one that lets teams tune the full loop: model, tools, memory, evals, policy, and runtime. Watch the metric nobody puts in the launch headline: cost per successful governed task, not cost per token. #AIAgents #EnterpriseAI #NVIDIA #LangChain #PlatformEngineering #MLOps

The cheapest model in the stack may become the most expensive governance decision. CNBC reports (cnbc.com/2026/07/08/chi…) that U.S. lawmakers are probing the growing use of Chinese AI models by American companies. The reason is easy to understand: these models are closing performance gaps while staying cheaper to use. CNBC names Cursor and Airbnb in the context of committee letters, and notes that Coinbase's Brian Armstrong and Lindy's Flo Crivello have publicly discussed using Chinese models to reduce costs. This is where the debate gets lazy. One camp says, "Ban them." Another says, "Use whatever is cheapest." Both skip the operating question that CIOs actually have to answer: what risk are we accepting, in which workflow, with what data, under what control plane? A model used for marketing copy is not the same as a model used for vulnerability discovery, claims review, trading support, or clinical operations. Origin matters, but workload matters too. Data path matters. Hosting matters. Auditability matters. The ability to prove what was sent where matters. The practical move is not a blanket panic. It is a model risk tiering system that procurement can actually enforce. If a cheaper model sits behind an approved U.S.-based provider, handles low-sensitivity tasks, has logging, and never touches regulated data, the tradeoff may be reasonable. If it touches source code, PHI, security workflows, or customer identity, the burden of proof should be much higher. The board question is simple: do we know where foreign-origin models are already embedded, or will we discover it from a regulator, a customer, or a congressional letter? #AI #EnterpriseSoftware #CIO #AIGovernance #RiskManagement #Cybersecurity

I’m not particularly thrilled about the upcoming frontier model capabilities, but I’m quite excited about the potential evolution of J-space in the next few major versions.

Google's AI infrastructure report (cloud.google.com/blog/products/…) says 83% of organizations need upgrades for agentic AI. That number sounds like cloud marketing until you look at the mechanism. A chatbot asks and answers. An agent plans, retrieves, calls tools, checks state, writes back, and sometimes loops through that process many times. One user request can turn into hundreds of downstream actions. That changes the infrastructure problem from "Can we call a model?" to "Can we govern a distributed execution system?" This is where many AI roadmaps are underbuilt. The hidden costs are not just tokens. They are egress, memory, latency, fragmented data, idle accelerators, duplicate pipelines, identity gaps, and audit trails that were never designed for autonomous software actors. Google says 62% of leaders are seeing a significant "inference tax" and 79% cite security, governance, and MLOps as the top challenge to scaling inference. The Day 2 implication is simple: agents make infrastructure political again. Cloud, data, security, compliance, and app teams can no longer optimize in isolation. The agent needs context from everywhere, permissions across systems, and a control plane that can explain what happened after the fact. The board-level question is not whether the company has an AI strategy. It is whether the infrastructure can handle autonomous action without creating an expensive, invisible mess. #AIInfrastructure #AgenticAI #CloudComputing #PlatformEngineering #MLOps #EnterpriseArchitecture

Ukraine's AI policy is a preview of where enterprise AI is going. Ukraine will favor AI systems it can run on its own servers, especially for government, business, and military use. The trigger was not academic. It followed U.S. restrictions that forced Anthropic to cut access to powerful models, and it reflects a broader fear: remote AI can be restricted, switched off, or governed by someone else's politics. This is the part of AI sovereignty that buyers should take seriously. It is not only about national pride. It is about operational continuity. If an AI system becomes embedded in public services, defense workflows, clinical operations, financial controls, or critical enterprise processes, then dependency on a remote-only model is a business risk. The question becomes less "Which model is smartest?" and more "Can we keep running when access changes?" A practical risk register for AI buyers: - Can the model run in our jurisdiction? - Can we self-host or use a controlled deployment? - What data leaves the environment? - What happens if the vendor changes policy? - What workloads require local fallback, even if the cloud model is better? The interesting part is that Ukraine's official said in a media report supplier nationality is not the deciding factor. Control is. If the vendor can support on-premise deployment, it can be considered. That is the right framing. In high-stakes AI, model quality matters. But control becomes part of quality. #AISovereignty #AIGovernance #EnterpriseAI #Cybersecurity #AIAgents #DigitalSovereignty

Healthcare software has a strange problem. Every company building for clinics/hospitals spends most of its money rebuilding the same foundation: compliance, EHR plumbing, audit trails, doctor sign-off workflows. Hikigai (hikigai.ai) built that foundation once, as a platform. It works like an autonomic nervous system and it lets any developer build apps that can transform healthcare from reactive to autonomic and from generic to personalized, without touching the hard parts. Proof it works: we built our own AI Scribe and CarePilot on it, both live with early paying customers in the US. Hikigai is raising funds under SAFE model (from Y-Combinator) and it will fund GTM for the apps and build an ecosystem of developers and ISVs around the platform. If you want to play a role supporting the transformation of healthcare at a global scale, let us talk.