Notes and essays from Tarry Singh's studio: AI strategy, deep-tech architectures, and the craft of drawing speculative work.
Gulf states have issued millions of AI certificates, yet telco rosters are not moving at the same pace. This essay examines the gap between what Saudi Arabia's SAMAI, the UAE's AI Nation Afaaq, and Qatar's operator programmes report as training headcounts and what shows up in actual hiring specifications and retention figures — and argues that the certificate is a promise the operator has not yet…
Germany's Energy Efficiency Act requires datacentres commissioned after July 2026 to redirect at least 10% of their waste heat, a rule public since 2023 that arrived before the offtaker ecosystem could support it. Operators now completing fit-out are engineering around the exemption clause rather than meeting the target, because the district heating networks the law assumes do not exist at the…
Most enterprises deploying AI agents cannot list what is running, who approved it, or what it costs: a governance gap that is today's operating environment, not a future risk. Agentic workflows trigger ten to twenty LLM calls per task, and token volumes have grown sixty-fold since 2023. The agents you cannot name are the ones that will fail an audit, breach a data policy, or exhaust months of…
Alphabet's Q2 2026 produced −$5.9 billion of free cash flow, the first negative quarter in the company's public history, as hyperscaler capital expenditure reached 31–83% of revenue, ratios previously confined to utilities and telecoms. That arithmetic has changed the CFO's job description: debt issuance is now a quarterly operational necessity rather than an opportunistic transaction, and capital…
A study published in Radiology found that general radiologists using an AI-driven workflow matched subspecialist cancer detection rates across nearly 578,000 scans. The result is real, but detection rate is not throughput, and the staffing crisis driving 2.5% annual attrition in radiology remains unmeasured. The clinical win and the capacity win require different budgets; the trial answered only…
Nigeria's 3MTT reports a 99.6% completion rate across 135,000 fellows, a figure measuring whether they finished an assessment, not whether they hold jobs that use the training. The IMF's baseline scenario projects a 0.2% productivity lift for sub-Saharan Africa over the coming decade under current conditions. Until programmes publish cohort-level employment outcomes at six and twelve months, the…
The EU AI Act's Article 12 logging requirement is an architectural obligation, not a monitoring one. Guardrails embedded inside the application produce audit evidence the application controls; guardrails deployed as a gateway produce records the application cannot reach. That distinction decides compliance, and organisations retrofitting the wrong architecture will find the fix costs three to five…
A ratified specification is not a commissioned cluster. UALink 2.0 promises 1,024-accelerator pods at half the TCO of NVSwitch, but evaluation hardware is not expected until late 2026, and your purchase order is due in August. The more productive question for enterprise finance is whether InfiniBand or Ethernet closes your deployment faster than a roadmap that has not yet touched silicon.
The Predetermined Change Control Plan is FDA's mechanism for letting AI-enabled medical devices update without a new submission. It is not a solution to the adaptive-AI problem. It is a scoping exercise written in advance. The modifications a vendor can ship without re-filing are precisely and only those the engineer defined eighteen months before knowing what the field would ask for.
Space42's decision to replace Chinese-made autonomous vehicles on Saadiyat and Yas Islands with a Korean fleet from Autonomous A2Z is the first legible signal that Gulf sovereign operators are treating AV supplier selection as a geopolitical-risk register entry. A foreign-controlled autonomous fleet can be paused on someone else's political calendar, as WeRide demonstrated during the Iranian…
India curtailed 8,133 gigawatt-hours of solar power in the second quarter of 2026, roughly 18% of average monthly generation. The forecasting models worked. Transmission infrastructure was not built in time to carry the output. The country's grid AI layer is functioning as designed; the ceiling it optimises below is set by copper, steel, and land-acquisition law.
The Federal Reserve placed agentic AI outside its model-risk framework in April; the European Commission issued transparency guidance two weeks before enforcement began in August. For banks operating across both jurisdictions, the collision is immediate and the governance gap is real. The first regulatory action will likely come for failing to document why deployment was safe at all, rather than…
On 2 August 2026, national market surveillance authorities across the EU began enforcing Article 4 of the AI Act, the staff AI literacy obligation that has technically been in force since February 2025. Most European universities have confused talking about AI with ensuring their staff can actually handle it. The first penalty will almost certainly arrive as an aggravating factor when a high-risk…
More than $130 billion in U.S. datacenter projects was delayed or abandoned in Q1 2026, and water arrived before power as the binding constraint. Across more than 20 states, moratoriums and permitting overhauls have frozen roughly a third of the new-build pipeline. Closed-loop and immersion cooling can solve the problem; the capex delta means only hyperscalers are currently able to absorb it.
Uber burned its entire 2026 AI coding budget by April, and it is not an outlier: unexpected AI costs changed a business decision at 62% of organisations this year. A charge only reaches directors once it has cleared the spending authority of whoever caused it. Budgets set in 2025, before agentic workloads ran at production scale, are now carrying orphaned agents that burn tokens with no owner and…
Enterprise AI training is a governance failure. Most firms are deploying tools faster than they build genuine competence, outsourcing upskilling to social media and self-directed learning while counting adoption rather than capability transfer. The evidence is unambiguous: fragmented, optional programmes produce anaemic ROI, a widening workforce fracture, and a skills hollow that no SaaS contract…
Two Gartner surveys published in May 2026 reveal a measurement crisis enterprises have not yet named: time saved per AI seat is real, but organisations lack the controls to redeploy that capacity or quality-gate its output before it reaches the P&L. Between 60% and 89% of firms see zero detectable productivity impact despite widespread adoption. That is a governance problem that registers as a…
The UAE has assembled the most complete sovereign healthcare-AI stack any country has built: sovereign compute, a million-sample national genome programme, and a clinical model benchmarked ahead of GPT-4. The model's own authors say it is not yet fit for clinical use. That caveat is the most precise evidence about where the Gulf investment earns its return: in the biobank and the data-sovereignty…
India is installing, at national scale, a curriculum-grounded retrieval-augmented tutor: CBSE mandating AI from Class 3 upward, SATHEE running in twelve Indian languages, a target of 1 million trained teachers by 2027. The retrieval architecture is the right one. What is missing from every deployment surveyed here is the teach-versus-solve evaluation loop: the independent, automated check that…
LineShine, which debuted atop the June 2026 TOP500, is the first system to exceed two exaflops of sustained double-precision performance without GPU accelerators. It was built under US export controls from indigenous Arm processors, on-package HBM, and a custom fat-tree fabric called LingQi. The ranking is the least of it: China has closed the interconnect gap that sovereign exascale programmes…
BMW's ten-month, 30,000-unit deployment of Figure AI's humanoid at Plant Spartanburg has shifted the industry's central question from 'can it work?' to 'at what cost, and where?' With unit costs projected below $17,000 by 2030 and current ROI timelines compressing toward 14 months, humanoid robots are becoming a capital-allocation decision. The binding constraint has moved from hardware to…
Four US banks (Bank of America, JPMorgan Chase, Goldman Sachs and Wells Fargo) switched on production-grade generative-AI systems across their contact centres in the same quarter that federal supervisors carved these technologies out of revised model-risk guidance. Held against the Toyota Production System, the rollout reveals an industrial line running at scale with no jidoka cord to pull and no…
UNEP's Methane Alert and Response System has processed over 1.3 million satellite observations since 2023, with AI models correctly flagging 80–85% of confirmed methane detections before expert review, helping mitigate an estimated 1.2 million tonnes. Yet 88% of MARS alerts still go unaddressed. The constraint is no longer algorithmic; it is the choice, made repeatedly by governments and…
SR 26-2, the Federal Reserve's updated model-risk guidance, carves generative and agentic AI out of supervisory scope. The carve-out is a placeholder while regulators decide what replaces SR 11-7. The EU AI Act's high-risk obligations arrive on 2 August. For multinational banks, the result is a structural supervision paradox: US headquarters has time but no clarity; European subsidiaries have…
The European Commission has published an AI literacy framework for schools but has not resolved the question that actually moves labour markets: which credentials count, against what standard, and who recognises them across member states. With EU AI Act enforcement beginning in August 2026, private vendors are filling the compliance vacuum faster than any public body can audit, and may end up…
Two Korean laws that took force within months of each other are not the same law, and reading them as one will trip on the seam. The revised Network Act names eight platforms, Naver, Kakao, Google and Meta among them, and requires half-yearly transparency reports on content moderation. The AI Basic Act separately demands visible labels on AI-generated advertisements. The result for ad buyers is…
An enterprise client burned $500 million on AI services in a single month, a consequence of deferred governance rather than reckless strategy. AI slop debt, the compounding cost of zombie agents, ungoverned prompts, and abandoned POCs that still bill, has crossed from engineering irritation to balance-sheet liability. When 73% of AI initiatives exceed budget, the board question is no longer which…
When WongPartnership signed with Harvey, it acquired a seat licence and a compounding stack with it: wrapper-maintenance costs running 30–60% above the sticker, prompt libraries with no clear owner, and a hallucination-liability perimeter across three jurisdictions that Vietnam's new AI statute just made more expensive. The Forrester 400% ROI ceiling belongs to firms with clean, documented…
Saudi Arabia's mandatory cession ladder reached 30% in January 2025; the UAE Central Bank's Guidance Note on responsible AI followed in February. Together they define what the AI story in Gulf insurance actually costs. A domestic reinsurance mandate sits on one side and a documented compliance overhead on the other: annual bias testing, Arabic-language disclosure, mandatory human review at the…
India's mobile networks are mid-transition to AI-managed control loops; the public response has been fifty certificates after a one-month course. The arithmetic does not close. Some 380,000 operational workers need meaningful retraining by 2030, and cohort-based convocations are theatre for the annual report. What will actually work is paid, on-shift reskilling anchored to a live network problem,…
China's agricultural labour force has halved since 1991, with three-quarters of what remains now over fifty. Muyuan Foods runs 583 commercial hogs per employee, a ratio only possible because sensor grids and vision models have replaced the stockmen demographic gravity removed. The cost improvement is real and auditable; immunity from the pork price cycle is not, and the Solow productivity paradox…
American ports are deploying AI on a two-to-three-year clock onto crane infrastructure that will take a decade to replace, roughly 80% of it still built by Shanghai Zhenhua. The sovereignty gap is the distance between those two timelines rather than a gap between competing models. Three failure points are already in the calendar: the ILA contract running to 2031, no credible domestic crane…
The question in agentic commerce has moved. A language model can assemble a cart, so the live issue is who holds legal authority when that cart signs itself against a live payments rail. The planner-executor split now carries cryptographic mandate obligations with EU regulatory teeth. The missing component in every production deployment studied here is an eval harness rigorous enough to catch the…
The WHO-approved chest X-ray AI models running on tablets in Turkana are working. What is not yet settled is where the stack underneath them actually lives: training pipelines, reference datasets, and GPU clusters that remain, for now, in foreign clouds. The sovereignty question in African clinical AI gets decided in procurement documents, on the clock that replaces the current generation of X-ray…
Satya Nadella's essay on the reverse information paradox names a real problem: enterprise buyers surrender proprietary knowledge to AI vendors with each query: prompts, corrections, evaluation harnesses. The framing is analytically sound. It is also written by the CEO most commercially exposed to a future in which that knowledge lives outside his cloud tenant. The enterprise buyer's job is to…
KEPCO's AI-driven grid management system and its MOU with OpenAI are genuine operational improvements. They address the wrong layer of Korea's energy problem. The country faces a structural generation deficit of roughly 24.7 GW to serve AI data centres and chip fabs, with transmission projects running late and debt at ₩202 trillion constraining capital spend. Software cannot conjure copper where…
Singapore's Monetary Authority has published the first regulator-authored document to name, at runtime, the exact control gates between an autonomous agent's proposed action and a live ledger. SAFR's policy engine, runtime validator and append-only audit ledger map onto what DBS and Grab Finance already run in production. For model risk officers from Jakarta to Frankfurt, this is the document to…
The EU-OECD AI Literacy Framework, released 18 June 2026, is the most coherent curriculum structure of its kind: four domains, 19 competences, offline adaptations for under-resourced schools. It does not credential, fund teacher upskilling, or close the 9.7-million ICT specialist gap the EU's Digital Decade requires by 2030. Article 4 of the AI Act has obligated staff training since February 2025.…
Across the Gulf's flagship industrial operators, public commitments to AI reskilling have assembled a recognisable two-track system: research master's programmes at KAUST and MBZUAI at the top, SDAIA and ADNOC bootcamps at the bottom. The publicly disclosed plan documents have no middle rung, no plant-embedded practitioner who can translate an anomaly-detection alert into a shutdown call at a DCS…
Nvidia's warm-water cooling eliminates water use inside the data centre and nothing beyond it. The bulk of AI's water footprint, roughly 54% by the IEA's accounting, sits behind the meter in fossil fuel plants supplying the grid when renewables fall short. Until disclosure extends to end-to-end water intensity per compute-hour inclusive of generation fuel source, the industry is optimising for a…
Uber exhausted its entire 2026 AI coding budget by April, four months into the year. Beneath that headline sits a three-layer debt stack: technical debt, ungoverned prompt sprawl, and cost overhang accrued during a subsidised-inference era now ending. Organisations that treat AI spend as an architectural concern from day one will survive what follows; those that audit costs only after the invoice…
Two years after Cabinet approval, the IndiaAI Mission has disbursed roughly Rs 400 crore of its Rs 10,371.92 crore five-year outlay. Ambition is not the constraint. The instruments to move public money into foundation-model builders and GPU tenders did not exist at announcement. The second half must close a machinery gap before the headline commitment is quietly trimmed in successive budget notes.
Baidu's Q1 2026 filing contains two businesses moving in opposite directions: AI Cloud revenue up 49%, online advertising down 21% year on year. This is a structural argument about what happens to search-driven ad economics once AI answers resolve queries before a sponsored result ever loads, rather than a Baidu-specific execution story. The ad line, this piece argues, will structurally undershoot…
Corporate AI mandates now outrun demonstrable capability. Google ties usage to performance reviews, Meta tracks AI-assisted output, Microsoft has declared it non-optional. 82% of organisations provide AI training; 59% still report a skills gap. Access is not the deficit. Completion certificates measure the wrong thing, and firms are applying one-time training models to a skill that demands…
Gartner surveyed employees in Q1 2026 and found that 19% reported no time saved with AI, while their employers went on tracking 'hours saved' as the primary success metric.
Brazil's sovereign-AI strategy is being built by a set of overlapping institutional decisions rather than by a chip fab or a frontier model: spectrum obligations attached to 700 MHz licences, a state-cloud intermediation layer at Serpro, Telefônica Brasil's portable OpenShift platform, and a CPQD-led public-services model. The coherence is architectural rather than planned, and the cell tower is…
Japan's hillside farms cover 40% of cultivated land, with fragmentary cell coverage and a workforce averaging 69. They force a reckoning with agent architecture that enterprise AI teams have deferred. When the link is intermittent, the planner-executor split is the only arrangement that lets the tractor stop safely. A physically independent critic completes the trio, and no board should approve an…
Microsoft is sitting on roughly $80 billion of Azure orders it cannot fill. It has the demand and it can buy the chips. What it cannot get is power. The binding constraint on AI has quietly moved from the fab to the substation, and the hyperscalers' answer, becoming their own utility, is rational for them and corrosive for everyone else standing in the same interconnection queue.
Gulf e-commerce fulfilment has quietly become a manufacturing problem: closed-loop robotic sortation, AI-orchestrated pick paths, and condition-based drive monitoring now define the floor at Noon's KEZAD centre and Amazon's DXB5 expansion. Whether the capex is paying back cannot be answered from press releases. The throughput-per-labour-hour and sortation-utilisation figures that would settle the…
India's largest generic pharmaceutical manufacturers now procure seventy to 85% of their electricity from renewables, a figure that flatters the underlying problem. Pharmaceutical cleanrooms account for fifty to 80% of a sterile plant's total energy and run continuously at setpoints regulators inspect without warning; changing the grid mix does not move the meter. Model predictive control of…