Noor Ejaz Chaudhry

My review of gender justice in 2021 for The News on Sunday. I talk about what went wrong, what must be celebrated and who must we attribute these victories to. Closing the gap thenews.com.pk/tns/detail/919…

Anthropic just got the most valuable credential in enterprise AI. It came from a Pentagon blacklist. The supply chain risk designation has historically been used for Chinese state-linked firms. Anthropic is the first American company to receive it. The reason: it refused to let Claude be used for autonomous weapons and mass surveillance. A federal judge wrote the government "appears designed to punish Anthropic." Seven competitors took the contracts. In a European bank's compliance review or a MENA sovereign fund's AI vendor evaluation, that story reads differently. Not as a loss. As a proof.

PitchBook titled their OpenAI analyst note this quarter "The IPO That Cannot Afford to Wait." The diagnosis is right. The reason behind it is more uncomfortable than the headline suggests. OpenAI is projecting $14 billion in losses in 2026 on roughly $25 billion in revenue. Annualized revenue grew from $6 billion at end of 2024 to $21.4 billion at year-end 2025 to $25 billion by February. Per CNBC's reporting this week, the company missed multiple revenue targets and growth estimates are falling short. The Amazon comparison keeps appearing in coverage. Amazon lost money for seven consecutive years. It worked. But the parallel breaks under pressure. Amazon's losses from 2001 to 2007 were funding physical infrastructure: fulfillment centers, servers, trucks. Each dollar of capex built capacity with calculable unit economics. The cost per shipment fell predictably as volume scaled. The moat was physical, defensible, and time-consuming to replicate. OpenAI's $14 billion in losses is funding compute and model weights. Both are replicable. Google, Microsoft, Meta, and Anthropic are all training frontier models at scale. The barrier to entry isn't capital anymore. The hyperscalers solved that. The barrier is distribution, and OpenAI doesn't control distribution the way Amazon controlled the checkout button. ChatGPT is the most-used consumer AI product in the world. That's real. But Microsoft is pushing Copilot into every Office installation. Google is pushing Gemini into every Chrome tab and Android keyboard. Meta is pushing AI assistants into every WhatsApp conversation. OpenAI's consumer distribution depends on users opening a separate tab. PitchBook is right that OpenAI cannot afford to wait. The IPO window exists while the ChatGPT brand is still synonymous with AI. The longer the wait, the more that synonymy gets distributed across five interfaces people already use every day. The $1 trillion valuation prices OpenAI as the default AI layer of the internet. Defaults live in distribution. And in distribution, OpenAI is competing against the companies that own the interfaces.

The only Mag 7 stock up double digits after earnings this week was the one whose earnings call contained the sentence: "revenue would have been higher if we could meet demand." That was Google Cloud. $20 billion in the first quarter. Growing 63% year over year. Running nearly $2 billion above analyst estimates. And still capacity-constrained. Per the Alphabet earnings release, Cloud operating income tripled in a year, from $2.2 billion to $6.6 billion. The remaining performance obligations backlog nearly doubled in three months to over $460 billion. That backlog is contracted, not projected. Those are customers who have signed deals and are waiting for Google to build the infrastructure to serve them. Alphabet stock climbed 10% Thursday. Year to date, it's up 23%, the best performance in the Mag 7 by a significant margin. Meanwhile, Meta raised its 2026 capital spending guidance by $10 billion, citing higher component costs. Its stock fell 6%. Microsoft reported $82.9 billion in first quarter revenue, beat analyst estimates, and slipped. Amazon reported $181.5 billion in revenue, beat estimates, and slipped. The market is not punishing earnings beats. It's punishing capex without proof. Meta, Microsoft, and Amazon each exceeded top-line estimates. Combined 2026 capital expenditure across the four hyperscalers is on track to exceed $650 billion, the largest single-year corporate capital commitment in history. The market looked at $650 billion and asked one question: where is the revenue return? Google answered. The others didn't. The mechanism matters. Google Cloud's 63% growth wasn't a new product launch or a market share grab. Enterprise AI infrastructure demand is outpacing the server capacity available to fulfill it. The constraint is supply, not demand. Every bear case on AI cloud adoption assumed the opposite: enterprises would be slow to buy. The actual limiting factor in Q1 was that Google couldn't build fast enough. There's a structural reason Google crossed the proof threshold first. It owns the full stack. TPU Ironwood, the seventh-generation custom chip, which SemiAnalysis pegs at 44% lower total cost of ownership than NVIDIA's GB200 in a finished rack. Gemini. Google Cloud Platform. The data, the distribution, the enterprise sales motion, the model. When Microsoft sells Azure AI, the chips come from NVIDIA. When Google sells GCP AI, the chips come from Google. The margin profile and the supply chain risk are entirely different. Anthropic locked in 1 million TPUs on Google infrastructure. Meta signed a multibillion-dollar TPU lease. Two of NVIDIA's largest customers are now also paying Google for AI infrastructure. Google is on both sides of the AI capital cycle. The other hyperscalers will build proof. Azure Foundry, AWS Bedrock, and Meta's internal AI are credible platforms. But the timing window isn't symmetric. Google's $460 billion contracted backlog is not a sales pipeline. Those customers have signed contracts and are in a queue. The revenue is already committed. In 2015, when AWS reported its first standalone financials, the headline was "Amazon is a cloud company that happens to sell boxes." Google's 2026 headline may be "Google is an AI infrastructure company that happens to have a search business." Cisco owned the internet pipes in 2000. ExxonMobil owned the energy price cycle from 2007 to 2011. Both lost the top spot within a decade because they owned one layer, not the stack. Google has the stack. The demand is queued. The only constraint is servers.

Circle's CPN Managed Payments is not a stablecoin adoption story. It is the moment correspondent banking became opt-out infrastructure. Launched April 8, 2026, CPN Managed Payments lets banks and payment service providers settle cross-border in USDC without holding a single dollar of digital assets. Circle absorbs minting, burning, FX exposure, and compliance. The client sees fiat in and fiat out. The stablecoin layer is architecturally invisible. This deserves more attention than it got. The banking sector spent the last five years blocking crypto adoption on three grounds: digital asset custody, licensing complexity, and compliance overhead. Circle's product eliminates all three. The bank never touches the asset class it spent five years lobbying against. The transition cost is zero. SWIFT's 11,000 member institutions processed roughly $5 trillion daily in 2024 but still settle on an end-of-business batch cycle with cutoff times and weekend gaps. CPN settles in seconds, 24/7, with no correspondent bank in the chain. Correspondent banks charge 25 to 40 basis points per transaction on average. On cross-border volume of that scale, that friction is measured in billions per day. The stablecoin market is not small context here. Supply topped $300 billion in 2025. Transaction volume hit $27.6 trillion in 2024, more than Visa and Mastercard combined. CPN is the distribution rail that brings that volume into banking's core settlement stack, not its crypto desk. Circle is not competing with Tether. Tether is a dollar proxy for emerging market treasuries. Circle is competing with SWIFT, and it is doing it by making the technology look like nothing at all. The stablecoin narrative is the wrapper. Correspondent banking disintermediation is the product.

Pakistan's $41 billion remittance projection is not a diaspora success story. It is the clearest available signal that the country's domestic economy has lost the productivity competition with its own citizens abroad. Pakistan's goods exports for FY2026 stand at approximately $32 billion. Remittances are projected at $41 to 42 billion. The gap is $9 billion and growing. In FY2024, remittances were $30.3 billion. In FY2025, $38.3 billion, a 26.6% surge. March 2026 alone hit $3.83 billion, a single-month record. The structural shift is more revealing than the absolute number. In 2000, remittances were 1.1% of Pakistan's GDP while goods exports were 9.1%. Those lines have now crossed: remittances at 9.3% of GDP, goods exports at 7.9%. Over 25 years, the country's most productive economic activity migrated abroad. Approximately 4.5 million Pakistani workers are in GCC countries. Saudi Arabia and the UAE account for 54% of total inflows. That concentration creates a structural exposure rarely stated directly: Pakistan's foreign exchange reserves depend on Gulf labor market conditions, not domestic economic policy. The policy implication is precise but uncomfortable. The remittance figure improves when workers abroad earn more, not when Pakistan's economy performs better. GDP growth and remittance growth are functionally decoupled. The country's most productive export is not a textile or a food product. It is a Pakistani engineer placed in a Gulf economy that pays 5 to 10 times the equivalent domestic wage. Pakistan has optimized for extraction. That is not a story that ends without structural reform at home.

Isaac Newton ran the numbers on the South Sea Bubble, saw it was a fraud, and bought in anyway. He lost £20,000. The math was never the problem. 🧵

At the peak of tulip mania, a single bulb sold for 10,000 guilders. A skilled craftsman earned 300 guilders a year. People sold their houses to buy flowers. 🧵

Capital discipline just became an AI strategy. Apple reported its March quarter this week. Record revenue. Services growing. Stock up 4% after-hours. Capital spending for the quarter totaled $1.9 billion. Down 36% from a year earlier. Compare that to the rest of Mag7. Microsoft guided $190 billion in 2026 capex. Amazon guided $200 billion. Meta guided $125 to $145 billion, then watched its stock fall 6% for the privilege. Alphabet guides near $180 billion. Four companies. One year. Roughly $650 to $700 billion in AI infrastructure spending. Meta spent more on AI infrastructure in a single quarter than Apple will spend in a full year. Meta's stock fell. Apple's stock rose. The consensus going into this earnings cycle was that capex is the primary AI signal. The companies deploying the most capital are making the boldest long-term bet. Every dollar of announced infrastructure spending lifted multiples through 2024 and into 2025. The market's memo changed this week. Meta beat on revenue. Meta raised its capex guide. Meta fell 6%. Microsoft grew AI revenue 123% year over year to a $37 billion annualized run rate. Microsoft slipped 2.5%. Alphabet grew Cloud revenue 63% to an $80 billion run rate. Alphabet traded higher. The dividing line was not size of spend. It was whether the revenue curve is visibly catching the capital curve. Apple's case is structurally different from the start. The $650 billion buildout is happening because cloud-side inference is expensive. Every query to GPT-5.5, Gemini Ultra, or Claude Opus hits a data center server. Those economics require massive upfront capital and a multi-year bet that inference demand scales fast enough to justify the depreciation cycle. Apple bets differently. The M4 chip inside an iPhone handles inference on-device. Writing tools, live transcription, photo processing, Siri Suggestions. None of that hits a data center. Apple's AI runs on hardware it already sold to 2.2 billion active devices. The inference bill went to the customer at purchase, not to a Virginia GPU cluster. No depreciation curve to explain to analysts. No GPU contract renewal. No quarterly capex guidance revision. The economics at the inference layer diverge sharply. Cloud inference means paying for every API call, every compute hour, every watt consumed. On-device inference means paying once, at hardware purchase, and running indefinitely on Apple Silicon. Apple's Services margin sits at 74% gross margin. The compute was already bought, already depreciated, already in 2.2 billion pockets. The historical parallel runs to the cloud transition of 2010 to 2015. Amazon, Microsoft, and Google spent aggressively to build out cloud infrastructure. Apple kept computing on-device and invested in silicon. A decade later, Apple ran the highest margin profile in the industry. 2.2 billion active devices. Services at 74%. Hardware that prints software economics. None of this means the hyperscaler bet is wrong. The $700 billion annual buildout may generate returns that dwarf Apple's profile over the next decade. AWS alone is at a $150 billion annualized revenue run rate. Azure AI is growing 123% year over year. That bet is working. But the market this week delivered a clear verdict. Capex no longer automatically buys multiple expansion. The cycle has moved from belief to proof. Apple was the only Mag7 name that never entered the proof round. The best AI strategy might be the one that never needed a data center.

The AI chip trade is sold as the diversified infrastructure bet. One company's missed growth target this week proved it functions as a single-stock proxy. The Wall Street Journal reported that OpenAI missed its own targets for revenue and user growth. ChatGPT reached 900 million weekly active users but fell short of the 1 billion goal the company set for end of 2025. Revenue slipped below internal projections as Gemini absorbed market share and Anthropic gained in enterprise coding. The market reaction exposed the actual structure of the trade. Nvidia dropped 3%. AMD dropped 3%. Broadcom dropped 4%. Oracle dropped 4%. One report about one company's internal targets moved the entire AI infrastructure complex in a single session. Oracle's move makes the concentration explicit. The company holds a $300 billion, five-year compute partnership with OpenAI, the largest single compute contract in corporate history. When OpenAI's CFO Sarah Friar reportedly warned colleagues the company might struggle to fund future compute agreements if revenue doesn't accelerate, Oracle was the name directly in the room. Nvidia's exposure is less direct but no less real. The hyperscalers spending $650 to $700 billion on AI infrastructure in 2026 are partly justifying that buildout through downstream inference demand, much of which flows through OpenAI. Less growth at the application layer means weaker pressure to order the next GPU generation. The diversification narrative holds over a long enough time horizon. Every enterprise, every geography, every new use case eventually needs chips. But the short-term volatility profile of the trade is a single-company proxy. Cisco sold gear to telecoms building for internet demand they couldn't yet see. When the ceiling became visible, Cisco fell not because of its own business but because its biggest customers' customers stopped ordering. The trade was right. The timeline was not. The AI chip trade may be right too. But it has one single-point-of-failure. And this week, it blinked.

Every piece of internet infrastructure was designed around one assumption: the thing at the other end is human. Cloudflare retired that assumption this week. Agents can now open a Cloudflare account, register a domain, start a paid subscription, and deploy code. Zero human steps after initial authorization. The Stripe payment limit defaults to $100 per provider per month. No dashboard. No API key copy-paste. No credit card form. The CAPTCHA asked: are you human? Cloudflare's new question is: which agent are you, who authorized you, and what are you allowed to spend? That is not a developer convenience. It is a rewrite of cloud infrastructure's identity layer. The question is not when agents take jobs. It is who controls the authentication and payment rails when agents become the majority of cloud's billing relationships.

Isaac Newton lost £20,000 in the South Sea Bubble. The man who calculated the orbit of planets could not calculate the madness of his neighbours. 🧵

Most people think of Tether as a stablecoin company. I did too. Then I started digging into how they’re deploying $13B+ in annual profit and realized it’s something much bigger. A thread on what Tether is actually building 🧵

I just started a Substack! You can subscribe to it here muneebnaseem.substack.com/?r=6oybp&utm_c…

Most people think of Tether as a stablecoin company. I did too. Then I started digging into how they’re deploying $13B+ in annual profit and realized it’s something much bigger. A thread on what Tether is actually building 🧵

Most people read today's core CPI print as a Fed story. I did too, at first. Then you factor in 9.1M b/d of Gulf production still offline and corridors that haven't absorbed the shortfall yet. The supply shock hasn't fully landed. The rate path is harder to read from here.

The White House just published research finding stablecoin yield poses limited systemic risk to bank deposits. Banks called the model flawed. That dispute is now the central CLARITY Act sticking point. Worth watching which Senate Banking members from deposit-heavy states break with the crypto consensus

From a payments-strategy lens, this is distribution beating brand: Tether backstopped Drift with $127.5M, conditional on migrating settlement from USDC to USDT. It's a market-share acquisition dressed as a rescue. Solana's stablecoin layer just went up for grabs

Something I've been tracking: Tether's 'People's Wallet' isn't just a product launch — it's a strategic pivot from B2B rails to consumer-direct. 500M users, $182B USDT in circulation. The question isn't whether stablecoins win payments. It's who controls the wallet interface.

Did anyone count how many tech players did Anthropic kill? Latest one is Figma…