Amazon Shows Its Hand and Microsoft Calls Copilot a Gimmick

Still Day One: Amazon on fire

Andy Jassy writes the most aggressive shareholder letter since Jeff Bezos's legendary “Day 1” manifestos. The Amazon CEO takes aim at Nvidia (“a new era has dawned”), taunts Intel (“98% of our top customers already use Graviton”), and also promises to break Starlink's satellite monopoly. The key message: Amazon's self-developed Trainium AI chips are already sold out for the next 18 months, and chip revenue has reached an annual run rate of $20 billion. If Amazon were a pure-play chipmaker, it would be at $50 billion—still far behind Nvidia's $216 billion, but with a steep growth curve. For context: Amazon is investing a record-breaking $200 billion in data centers in 2026, more than any other tech giant. OpenAI alone has pledged to spend $100 billion on AWS. → Tech Brew

Synthszr Take: Jassy is staging classic platform judo here: He's using AWS's own dominance as leverage against chip suppliers. We know this pattern from retail history, where Walmart and Aldi positioned their private labels against brand-name manufacturers. The difference: Amazon, with AWS, already controls 31% of the global cloud infrastructure and can offer customers an integrated package of computing power, software, and now its own chips. Trainium may technically lag behind Nvidia's H100, but if the chip is already optimized to run on the AWS stack and is 40% cheaper, many customers will accept playing second fiddle. The $200 billion bet is risky (the stock has already lost 15%), but Jassy is playing the long game: Whoever controls the infrastructure eventually dictates the terms.

Microsoft Copilot Just Wants to Play

Microsoft has hidden a remarkable sentence in its terms of service for Copilot: “Copilot is for entertainment purposes only.” The company explicitly warns against relying on the tool for important decisions, while simultaneously pushing the AI assistant into every corner of Windows—from Paint and Notepad to productivity applications. A Microsoft spokesperson described the wording as “legacy language” from when Copilot was still a Bing feature and announced it would be revised. The competition is acting similarly: xAI warns of hallucinations and offensive content in its terms of use, while Anthropic and OpenAI are preparing for multi-billion dollar IPOs. At Amazon, AI-generated code errors led to major outages, prompting management to order that senior developers must approve all code created by junior colleagues with AI. → futurism.com

Synthszr Take: Microsoft is playing the same game here as tobacco companies in the 1960s: aggressively marketing the product while legally covering themselves with fine print. The difference is that cigarettes at least consistently perform their function. The 'entertainment only' clause is not an oversight but a calculated risk mitigation for the day a Copilot error causes real damage. Amazon has already drawn its conclusions and is introducing a two-tier system for code: AI-assisted automatically means inferior until a human with seniority gives it the stamp of approval. This is the true fragmentation of the AI industry: not between expensive and cheap models, but between companies that take their own products seriously and those that sell them as entertainment.

Perplexity: The Agent Bet Yields 50% Revenue Growth

Perplexity has radically changed its strategy: Instead of continuing to compete with Google as an AI search engine, the company is now focusing on autonomous AI agents. Revenue increased by 50 percent within a month, and annual recurring revenue (ARR) is over $450 million. With over 100 million monthly active users and tens of thousands of enterprise customers, Perplexity monetizes through subscriptions ranging from $20 to $200 per month. The new product line, “Computer,” performs complex tasks with minimal human supervision, such as preparing tax returns with direct access to current IRS materials. Compared to the giants, Perplexity remains small: Perplexity reaches $2 billion in ARR, Anthropic $30 billion, OpenAI $20 billion. → TLDR AI

Synthszr Take: Perplexity proves that sometimes the best defense is a good offense. The company never really stood a chance against Google's search engine monopoly, so it simply redefined the market: agents instead of answers. The strategy is reminiscent of Netflix's shift from DVD mailers to streaming or Nvidia's transformation from a gaming chip maker to an AI infrastructure giant. The key lies in the pricing: While ChatGPT and Claude compete around the $20 mark, Perplexity creates a justification for ten times that amount with specialized agents. The tax AI is just the beginning; once users get used to software filling out forms autonomously, they will expect the same for insurance applications, government paperwork, and contract negotiations. Perplexity is betting that the market for AI assistants is already saturated, while the market for digital employees is just emerging.

Iran War (I): An Opportunity for Chinese Cloud Providers

Iranian drones hit three Amazon Web Services data centers in the United Arab Emirates and Bahrain on March 1—the first confirmed military attack on a hyperscale cloud provider. The attacks crippled banks, fintech platforms, and ride-hailing services in the Gulf. Iran has since threatened to continue attacking American tech infrastructure, including the $30 billion Stargate AI data center with Nvidia GPU clusters and proprietary OpenAI systems. Huawei Cloud is exploiting the uncertainty: “Single-region dependency is a thing of the past,” the company posted on X. “With uncertainties all around, multi-cloud is no longer optional — it's essential.” The Gulf states, which have invested heavily in American cloud providers since 2019 (Amazon and Microsoft each operate at least two cloud regions there), may reconsider their strategy. Although President Trump secured hundreds of billions of dollars in cloud and AI deals in Saudi Arabia, Qatar, and the UAE in May 2025, the military vulnerability of the data centers is changing the risk calculation. → The Download from MIT Technology Review

Synthszr Take: Huawei is playing the classic role of the third party who benefits, like Sweden in the Thirty Years' War. The Iranian attacks on AWS data centers transform cloud infrastructure from a technical risk to a geopolitical one: anyone who parks their data with American providers becomes a target in a conflict in which they are not even involved. The Gulf states find themselves in a dilemma reminiscent of the Non-Aligned Movement of the Cold War: they need state-of-the-art technology (only the US delivers Nvidia clusters on a relevant scale) but do not want to be crushed between Washington and Tehran. Huawei's 'multi-cloud' sermon is less a technical recommendation than a political offer: distribute your risks, and by the way, you get a provider that Iran won't bomb. The real question is whether the Gulf monarchies are willing to trade their technological sovereignty for military security.

Iran War (II): Information Fragments Outpace Fact-Checking Speed

In late March, Iran circulated a shaky video allegedly showing a U.S. F/A-18 fighter jet under attack. Iranian officials claimed to have destroyed the aircraft, which the Pentagon denied. The New York Times reconstructed how Iran turned a single post into a global audience of millions within 69 minutes. At 1:04 PM, an obscure X account with ties to Iran posted the video; one minute later, the Islamic Revolutionary Guard Corps followed on Telegram. Iranian embassies amplified the claim on X, and Iranian state television and Russia's RT picked it up within minutes. Pro-Russian influencers like “Megatron” reached nearly two million views, while American influencers like Ed Krassenstein (one million followers) further shared the unconfirmed report. Bot accounts mingled with real profiles, and short, affirmative comments with celebratory emojis flooded the comment sections. → Casey Newton

Synthszr Take: Iran has copied the principle of information warfare from the military doctrine of 'distributed lethality': instead of one massive propaganda strike, it relies on coordinated micro-attacks across multiple channels. The 69-minute cascade works like a DDoS attack on truth-finding—while fact-checkers are still analyzing the video's metadata, the claim has already gone through three rounds of amplification. The insidious part: even critical influencers become unwilling amplifiers because 'this is unconfirmed' looks the same in the feed as 'F-18 shot down.' The fragmentation of the AI industry is mirrored here in the fragmentation of the information landscape: just as open-source models undercut the frontier labs, cheap disinformation campaigns undercut the platforms' expensive verification systems. The next war won't be decided with F/A-18s, but by who can construct reality faster than the other can deconstruct it.

HappyHorse vs. Synthesia: The Mysterious Challenger

A mysterious video AI model called HappyHorse-1.0 has dethroned the established Synthesia and now leads the AA Ranking List (artificialanalysis.ai). The special thing is: no one knows who is behind it. The website shows no information about the developers; the model is open-source, has only 15 billion parameters, and generates videos in about a minute. HappyHorse masters synchronous ambient sounds (steps on ice, basketball sounds) and can directly generate speech in seven languages, including Mandarin, English, and German. The pricing is roughly the same as Seedance 2.0, so it's not below the market average. Many suspect an orchestrated marketing campaign behind its sudden appearance, including 'leaked' WeChat messages containing critical remarks about the competitor Jimeng. → AI Secret

Synthszr Take: HappyHorse follows the classic Silicon Valley playbook: anonymity as a feature, not a bug. It's reminiscent of Satoshi Nakamoto and Bitcoin, where identity-lessness became the brand identity. 15 billion parameters are tiny compared to GPT-4's suspected trillion, but apparently sufficient for video generation. The integration of synchronous audio shows where the real innovation lies: not in image quality, but in multimodality. The fact that an open-source model is beating proprietary competition fits the current pattern of AI fragmentation. HappyHorse is probably not a David versus Goliath, but a Trojan horse from a major player looking to test the market.

Fluency is Just a Simulation of Intelligence

The Business Engineer hits a raw nerve regarding AI usage: most people just produce wrong results faster with artificial intelligence. They create more text, more slides, more analyses, with greater speed and increased confidence, but the underlying conclusions remain unchecked. The problem isn't the technology itself, but the confusion between linguistic fluency and actual thinking. AI models are eloquent; the thoughts behind them often are not. This fluency, according to the author, perfectly camouflages mediocrity. His new book therefore addresses not the question of how to use AI, but how to think well enough so that AI is more than a sophisticated autocomplete for existing biases. → The Business Engineer

Synthszr Take: The Business Engineer describes a phenomenon already known to the Sophists in ancient Athens: rhetoric without substance sells better than uncomfortable truths. AI amplifies this old human weakness to an industrial scale, turning everyone into an eloquent charlatan who hides their intellectual laziness behind machine-generated prose. This is reminiscent of Gresham's Law from monetary theory: bad money drives out good if both have the same face value. In the AI era, fluent texts displace well-thought-out arguments because both appear equally competent on the surface. The real disruption is not that machines are learning to think, but that humans are stopping to think because the machine sounds so convincing.

Every Brand is Becoming a Content Factory

Tom Orbach's newsletter 'Marketing Ideas' documents a trend that is revolutionizing corporate communications: over 50 billion-dollar brands now operate their own media channels. OpenAI recently paid 'low hundreds of millions' for the 17-month-old podcast show TBPN. Plaid ($8 billion valuation) bought 'This Week in Fintech' with 200,000 subscribers. HubSpot acquired The Hustle (2.5 million subscribers) and Starter Story (800,000), Semrush snapped up Search Engine Land and Backlinko, while Zapier took over MakerPad. Orbach, who turned a podcast at Wiz into the #1 in cloud security and grew his own newsletter to 82,000 subscribers, promises a guide to building a 'Media Empire' on a zero-dollar budget. His thesis: in two to three years, having your own podcast or newsletter will be as standard for companies as having a website is today. → Tom's Marketing Ideas

Synthszr Take: This wave of acquisitions shows how companies are replacing their distribution costs with trust capital. What used to be called CAC (Customer Acquisition Cost) is becoming Audience Acquisition Cost—except that the multiplication for media brands is exponential. OpenAI's purchase of TBPN for 'low hundreds of millions' after 17 months corresponds to a monthly valuation increase of about $10 million, faster than any SaaS product could scale. The real leverage lies in reversing the funnel logic: instead of chasing customers through paid channels, they become voluntary subscribers who self-qualify. However, Orbach's zero-dollar thesis only works with a hidden asset: the time of top people who produce content instead of building products—a trade-off that only companies with enough leeway can afford.

NotebookLM Was Just the Beginning: Podcast AI Becomes a Modular System

Google's NotebookLM has struck a nerve with its Audio Overview feature: it turns any document into a podcast dialogue between two AI hosts. Now, an open-source project on GitHub shows where this is heading. Developer zarazhangrui has broken down the NotebookLM workflow into its individual parts and rebuilt it as a modular system. Users control everything: scripts, prompts, the hosts' voices, their roles, and their personalities. The tool converts any content—URLs, texts, files—into MP3s and feeds them directly into existing podcast apps like Apple Podcasts or Spotify. The clever part: you can have yourself analyzed by feeding it personal documents like resumes or meeting minutes and listening to two AI voices discuss your own thought patterns. → TAAFT - There's An AI For That

Synthszr Take: NotebookLM wrote the playbook, and now everyone is tearing out the pages and pasting them back together differently. What Google sells as a closed feature becomes a Lego set in the open-source community: the same TTS APIs, the same LLM prompts, but suddenly anyone can build their own podcast machine. This is reminiscent of the early days of WordPress when suddenly everyone could cobble together their own CMS. The most interesting aspect is the self-analysis function: you feed the AI your own digital footprint and then listen to two voices discuss your own blind spots. It's therapeutic voyeurism through the back door. It's the Synthszr podcast as a DIY project.