Here's what's been trending:
Opus 4.8 ships as Anthropic's "most honest" model: Same price as 4.7, agentic coding jumps to 69.2%, and it's 4x less likely to let code flaws slide.
Two trillion-dollar memory twins in one week: SK Hynix and Micron both crossed a $1T market cap as the AI memory supercycle went parabolic.
The Steam Deck nearly doubled in price: Valve hiked it from $399 to $789, the clearest consumer signal yet that RAMageddon has arrived.
Apple is betting on borrowed brains: Reports say it'll run a distilled Gemini model on-device using 15 years of custom silicon, ahead of WWDC on June 8.
China is locking its AI talent in: Beijing imposed overseas travel curbs on top researchers at Alibaba and DeepSeek.
ByteDance is weighing up to $70B in capex: The TikTok owner is also building its own CPUs as China's AI machine goes fully vertical.
Startup Intros has a simple mission: to help tech founders navigate everything in the early-stage, from compliance to capital.
🔥 Startup Intros Events Coming Up
On June 3, we are hosting the Anti-Shark Tank Pitch Night at Room 52 in NYC! There will be roasting of founders by the anti-sharks. Register on Partiful to attend!
Submit your app to pitch here!
May 29: Caffeine & Capital: Our morning coffee meetup is back at Corgi Cafe in FiDi! Grab coffee, meet founders and investors & register on Luma!
🤖 Opus 4.8 Lands as Anthropic Eyes Its First Profit
Anthropic shipped its most capable model the same week its first profit turned into a public argument. Opus 4.8 tops the coding leaderboards and pitches itself as the model that finally stops making things up. The catch is the $559M operating profit underneath the announcement, a number that depends on a lease Elon Musk and SpaceX's own S-1 describe two completely different ways.
🧠 The "Most Honest" Model Yet
Anthropic launched Opus 4.8 at the same price as 4.7, pitching it as more likely to flag uncertainties and less likely to make unsupported claims. The benchmarks back the framing: SWE-Bench Pro climbs from 64.3% to 69.2% (ahead of GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2%), SWE-bench Verified hits 88.6%, reasoning with tools goes 54.7% to 57.9%, and the model is roughly 4x less likely to let flaws in its own code pass unremarked. It also debuts Dynamic Workflows, a research-preview system that plans, fans out to hundreds of parallel subagents, then verifies, with the Claude Mythos security agent slated for the coming weeks.
The timing is pointed. Just last week, a new DeepSWE benchmark caught Opus 4.7 "exploiting a benchmark loophole" while GPT-5.5 took that crown. A model whose flagship feature is admitting what it doesn't know is the most direct answer Anthropic could ship.
💸 The $559M Mirage
Then came the headline number. Anthropic projected its first-ever operating profit, roughly $559M on $10.9B in Q2 revenue, up 130% from $4.8B the prior quarter, per investor documents the WSJ reviewed. The mechanism is efficiency, not just growth: the cost to serve a dollar of revenue fell from about 71¢ in Q1 to 56¢ in Q2 as inference got cheaper.
Read the same documents further, and the profit gets thinner. They project a roughly $14B loss for full-year 2026; Anthropic itself warned the profit won't hold once late-2026 infrastructure spend lands; the company has committed over $1T to compute and doesn't expect positive free cash flow until 2029. Critic Ed Zitron calls the framing a "Profitability Swindle." One profitable quarter, pulled from a year still set to lose billions.
⚡ The Colossus Asterisk
The clearest crack is the lease itself. Musk says the SpaceX Colossus deal is "a 180-day lease with 90-day notice," a short-term arrangement he insists SpaceX requested. SpaceX's own S-1 tells a different story: Anthropic agreed to pay roughly $1.25B a month, about $15B a year, through May 2029. One of those is a quarter-long handshake; the other is a three-year, ~$45B obligation. That gap sits directly under the profit line Anthropic just booked.
Translation: Anthropic is shipping the most self-aware model on the market while booking a profit on a year that still projects a $14B loss and a cost base its own counterparty describes two completely different ways. The model learned to flag its uncertainty this week. Would you book a profit on a lease your landlord says expires in six months?
🧠💾 RAMageddon: Memory's $1 Trillion Twins
Two memory makers crossed a trillion dollars in market cap in the same week, and a beloved $399 gaming handheld nearly doubled in price. The AI buildout stopped being an abstraction on a chart and started showing up on price tags.
📈 The Trillion-Dollar Club Has Two New Members
SK Hynix topped $1T in market value after a 9% jump, becoming the third Asian company to hit the milestone, with shares up more than 1,000% in a year. Days earlier, Micron hit $1T for the first time, closing up 19.29% after passing $700B earlier in the month. High-bandwidth memory, not logic, is suddenly the scarcest thing in AI.
🎮 The Steam Deck Tell
Consumers are footing the bill. Valve hiked the Steam Deck OLED from $549 to $789 for the 512GB model, citing "rising memory and storage costs." The Verge called the jump from the original $399 in 2022 to $789 today the end of an era for gaming handhelds. When the AI memory crunch reaches the checkout page of a gaming device, the supercycle is no longer a B2B story.
🌊 The Whole Stack Is Squeezed
The strain runs end to end:
Sivers: The Swedish optical-component maker is up ~1,700% YTD to a ~$2.5B cap, now one of the country's most-shorted stocks.
The Philadelphia Semiconductor Index: Up ~75% YTD, on track for its biggest annual gain since 1999.
Xiaomi: Posted its first quarterly revenue decline in three years, with net income down 57%, blaming the global memory price jump.
Optical supply: Shortages now span lasers, substrates, fibers, and connectors, per Nikkei.
Taiwan's tech firms: Completed a record $14.5B of debt deals this year, racing to finance AI capacity.
Translation: The cheapest input in computing for two decades just became the bottleneck, and the market is paying trillion-dollar premiums for whoever controls it. If your product roadmap assumed that memory and storage prices would keep falling, that assumption just broke. What does your unit economics look like at 2026 RAM prices?
🍎 Apple's WWDC Hail Mary
Apple is about to argue that the company most behind on AI is actually best positioned for it. Two weeks out from WWDC, the leaks suggest a strategy built on borrowing someone else's model and running it on its own silicon.
🤝 Borrowing Google's Brain
At WWDC, Apple is likely to showcase how 15 years of designing custom chips gives it an edge in local AI, reportedly using a distilled Gemini model that runs on-device instead of in the cloud. It's a clever reframe of a weakness: Apple can't build a frontier model, so it's selling privacy and on-device speed powered by a competitor's brain.
📱 Siri, Again
The consumer face of the push is a long-overdue Siri overhaul. Mark Gurman details a new UI, a chatbot-style app, and other major iOS 27 changes landing at WWDC on June 8. Apple has promised a smarter Siri before and missed; this time the demo arrives against a market that has already moved on to agents.
Translation: Apple is betting distribution and silicon beat raw model quality, and that running a good-enough model privately on a billion devices wins over a great model in the cloud. For founders, on-device AI just got a platform-level champion. Is "private and local" a real moat, or the consolation prize for losing the model race?
🇨🇳 China Pulls Up the Drawbridge
China spent the week treating AI like a strategic resource to be guarded, not just built. Beijing moved to keep its best researchers home, its biggest platform is going vertical on silicon, and its cheapest model just got cheaper.
✈️ Talent Can't Leave
Chinese government agencies began imposing overseas travel restrictions on individuals doing advanced AI work, now reaching private firms including Alibaba and DeepSeek. When a country starts controlling the movement of its engineers, it's treating model weights and the people who make them as matters of national security.
🏭 Building Its Own Stack
ByteDance is the clearest tell. The company is developing its own CPUs to support its AI infrastructure amid chip shortages, while weighing up to $70B of 2026 capex underwritten by its ~$50B in 2025 profit. To keep its researchers from defecting, it's even offering low-priced stock options tied to growth in its Seed AI division, a first for the company.
💸 Price War as Weapon
On the model side, the strategy is deflation. DeepSeek is making permanent a 75% discount on its flagship, dropping V4 Pro to $0.435 per million input tokens. While US labs raise at $900B valuations, China is racing inference costs toward zero.
Translation: China is building a self-contained AI stack, from custom CPUs to subsidized tokens to a talent base that can't board a plane. For founders, the cheapest capable models on earth now come with geopolitical strings attached. If your margins depend on DeepSeek's pricing, who actually controls your cost base?
🚀 Product Launch Quick Hits
Robinhood lets AI agents trade stocks: Users can connect AI agents to a dedicated account and virtual Gold Card to automate trades and credit card purchases.
Meta launches Plus subscriptions globally: Instagram, Facebook, and WhatsApp get paid tiers, plus tests of $7.99 and $19.99/month Meta AI plans and a $49.99/month creator plan.
ElevenLabs ships Music v2: The model can switch genres mid-track and handle complex vocals, built on licensed data and cleared for commercial use.
Google Docs Live: An AI voice tool for drafting documents rolls out this summer to AI Pro and Ultra subscribers.
Amazon opens Alexa for Shopping: The architecture, starter code, and learnings go to third-party retailers, starting with Kate Spade.
Oura Ring 5: A 40% smaller form factor with improved sensing ships June 4 for $399, up from the Ring 4's $349.
YouTube adds a custom feed: Signed-in US users can type a prompt to generate a constantly refreshed home-page feed.
Biohub's protein world model: The Zuckerberg-Chan institute releases a model for protein prediction, design, and discovery to researchers.
Qualcomm unveils the Snapdragon C: An entry-level ARM SoC for $300 Windows 11 devices, shipping in 2026 to take on the MacBook Neo.
SoFi launches SoFiUSD: A dollar-pegged stablecoin on Ethereum and Solana, now available to its 14.7M members inside the app.
Nvidia's Vera Rubin CPU benchmarked: Initial tests show the 88-core Olympus chip beating Intel's and AMD's x86_64 CPUs.
Nvidia retires the GeForce Control Panel: After 20 years, the legacy app is gone, with features ported to the Nvidia app.
👔 Personnel Quick Hits
Dropbox CEO Drew Houston steps down: After 19 years, Houston becomes executive chairman, with Dropbox Core GM Ashraf Alkarmi taking over.
OpenAI hires Colin Fleming as CMO: The ServiceNow CMO will lead business-unit marketing, succeeding Kate Rouch.
Wix cuts 1,000 jobs (20%): The cuts follow weak Q1 earnings and a ~50% stock collapse in 2026, citing the fast evolution of AI.
Synopsys gives Elliott a board seat: The chip-design software maker settles with the activist, adding Elliott's Jesse Cohn after ~two months of talks.
Illinois passes AI safety audit bill SB 315: The bill requires annual independent third-party safety audits of leading AI companies and heads to the governor.
CNN sues Perplexity: The filing alleges unlawful copying and distribution of CNN content after the two failed to agree on terms in 2025.
🌟 Editor's Note
At Startup Intros, our mission is to bring the latest founder-investor news straight to your inbox, keeping you ahead in the fast-paced world of Silicon Valley.
💭 Parting Thoughts
What a week: two memory makers worth $1T apiece, a Steam Deck that costs $789, Opus 4.8 promising to lie less, and a $1.25B-a-month lease nobody can describe the same way twice.
The through-line is scarcity finally hitting the income statement. For two years, AI demand was a stock-chart story. This week, it became a margin story: memory prices on consumer hardware, China hoarding both chips and people, Apple admitting it has to borrow a model to compete.
Build like compute is expensive again, because for the first time in a while, it seems like it is.
Till next time!
 | Dev Chandra |
Startup Intros Trending Thursday: Your trusted source for founder-investor insights, delivered with clarity and focus.