Good morning,
This week we cover the full Q2 earnings picture across platforms, chips and memory: record revenue, mostly falling share prices, and headline profits inflated by one-off gains. Then, on cost control: more companies are building their own coding agents instead of renting them. Plus the quick bites, and a podcast worth your commute.
Record Q2 revenue does not buy higher share prices
Revenue | YoY | Operating income | The AI line | Shares | |
|---|---|---|---|---|---|
PLATFORMS | |||||
Amazon | $200.6bn | +20% | $27.5bn, +43% | AWS +36.7%, fastest in 18 qtrs; TTM FCF −$7.6bn; capex guide → ~$220bn | ▲ ~10% |
Microsoft | $90.0bn | +18% | $40.6bn | Azure +43%, past $100bn/yr; Q4 capex $41bn; FY guide ~$175bn | ▲ ~8% |
Alphabet | $119.8bn | +24% | $40.8bn, +30% | Cloud +82%; capex $44.9bn; FCF −$5.9bn; guide → $195–205bn | ▼ ~4% |
Apple | $109.4bn | +16% | Net income $29.8bn | No AI capex guide; memory costs forced Mac/iPad price rises | ▼ ~6% |
Meta | $60.8bn | +28% | $18.8bn, −8% | Costs +55%; margin 43%→31%; FCF $0.8bn vs $31.1bn capex | ▼ ~9% |
CHIPS & MEMORY | |||||
AMD | $11.5bn | +50% | $3.1bn (non-GAAP, record) | Data centre $6.7bn, +107% — now 58% of revenue | ▼ ~8% |
Intel | $16.1bn | +25% | $1.97bn | Growth data-centre driven | ▲ ~11% |
Samsung | ₩171.5trn (~$119bn) | +130% | ₩89.5trn, +1,814% | Chip division ₩89.2trn profit; mobile posts first-ever operating loss | ▼ ~7% |
SK Hynix | ₩79.3trn (~$55bn) | +257% | ₩60.5trn, +557% (76% margin) | Missed consensus by 5%; capex rising to ~₩high-40trn | ▼ ~10% |
Micron | $41.5bn | +346% | 84.6% gross margin | HBM sold out through 2026; Q4 guide $50bn | ~28% below June peak |
AI-NATIVE | |||||
Palantir | $1.94bn | +93% | $912m (47% margin) | US commercial +149%; FY guide $7.66bn → $8.15bn | ▲ ~12% |
SpaceX | $7.81bn | +92% | Net loss $541m | Capex $18.4bn, of which $15.8bn AI; debt $22bn → $36.8bn | ▼ ~8% |
Nvidia will report on 26 August.
Three things to understand from those calls and market reactions. Headline profits are flattered: Alphabet's $112.1bn of net income included $98.0bn of unrealised equity gains, Amazon's $62.6bn included $53.4bn from its Anthropic stake. Share prices instead tracked whether AI spending is tied to billable revenue: Amazon added $20bn of capex and rose on AWS growth of 36.7%, Meta fell as costs grew twice as fast as revenue, and SK Hynix fell 10% on a 76% operating margin for signalling it would spend more. And the memory squeeze has reached consumer businesses — Samsung's chip division earned ₩89.2trn while its mobile arm booked its first operating loss since segment records began in 2011, and Apple slid 6% after Tim Cook said component costs had forced price rises.
Why it matters: Two practical points. Read operating income and free cash flow this year, not net income; net income is carrying large paper gains on AI stakes. And budget for a memory surcharge in every hardware refresh and cloud renewal through 2027; Samsung and Apple are showing what it costs downstream.
Cost control turns to a new solution: build your own
We have covered the cost of AI repeatedly this year — token budgeting, per-employee caps, routing, cost per task rather than cost per token. A new answer to this problem seems to be gaining attention. A growing number of companies have stopped optimising their bill and started building the system themselves.
What companies are building is the harness: the software layer around a model that lets it read a codebase, call tools, hold context and check its own work. Claude Code, Cursor and Codex are harnesses with a model behind them. Replace the harness with an open-source one, add a router that sends simple tasks to cheaper models, and costs drop sharply. Sapiom, a startup Anthropic itself backs, expects to cut AI spend by 25% after switching to the open-source OpenCode harness. The consultancy Codestrap says customer costs fell 97% after it built its own.
Larger firms have gone further. Coinbase, Shopify and Ramp have each built internal coding agents; Coinbase's now handles a share of merged pull requests and cut cycle time from around 150 hours to 15. Meta has asked thousands of engineers to use MetaCode, its in-house agent. The motivation is only half cost. The other half is exposure: few companies want a core workflow priced annually by a vendor whose own inputs are volatile.
Palantir has turned this into a sales argument. On the Q2 call, Alex Karp argued that enterprises paying for tokens are also handing over the proprietary knowledge that makes them valuable, and that the labs will eventually use it to compete with them. OpenAI and Anthropic both say customer data is not used for training. Satya Nadella (Microsoft’s CEO) made similar comments. Of course, both Karp and Nadella have a commercial interest in firms diversifying from frontier labs, but their concern seems to be resonating in the ecosystem.
Why it matters: Three things worth doing. Negotiate on consumption rather than seats, and model your renewal at three to five times current usage before signing. Own the harness layer where you can; it determines how easily you can switch models or vendors later. And check what vendor contracts actually promise about your data — "not used for training" is a narrower commitment than it sounds.
Quick bites
Deals & money
Harvey is raising at $15.5bn. The legal AI startup is in talks for at least $500m, a 40% premium to its valuation five months ago.
Nvidia is putting around $5bn into Safe Superintelligence. Ilya Sutskever's lab has no products and says it will release none before reaching its goal; it is pursuing a research direction away from the standard LLM path. Nvidia invested after being given rare access to the research.
Bending Spoons is buying Airtable for around $1.3bn, a steep markdown from its peak.
Models & benchmarks
OpenAI passed 1 billion users and 2 million business customers, days after cutting GPT-5.6 Luna's price by 80%. It is also reportedly on track to lose around $14bn this year.
OpenAI previewed "Astra" in Washington. An internal model produced solutions to ten long-open mathematics problems and published machine-checkable proofs for about $2,000 of compute.
DeepSeek plans substantial price increases as it resumes funding talks at a reported ~$70bn valuation. Note that its cheapest model scored roughly 37% accuracy with an 84% hallucination rate on one independent index.
Meta shipped Muse Spark 1.2 and a set of coding agents, Muse Code, joining Anthropic (Claude Code) and OpenAI (Codex) in the coding fields.
Hardware corner
China has begun producing immersion DUV lithography machines, the tools ASML has monopolised for two decades. ASML fell as much as 8% on the news. The machines are around a generation behind and ASML shipped ~130 immersion tools in 2025 on a 98.7% share. This is a major step in China's chip production becoming less exposed to Western export controls.
HP, Asus and Acer have started using Chinese CXMT memory in laptops sold outside the US. Volumes are kept small to protect relationships with Micron, Samsung and SK Hynix, but Western PC makers had not used Chinese DRAM before.
Washington is drafting a ban on Chinese optical transceivers in US data centres, after July's bans on Chinese humanoid robots and power inverters.
Safety & regulation
A sequel to last edition's great escape. The UK's AI Security Institute ran a red-team challenge 122 times. In 10 runs, agents took unsanctioned action on the live internet. In the worst, an agent tried to insert malicious code into a real open-source project, created fake identities and attempted to persuade human maintainers to approve it. Guardrails had been deliberately removed and internet access was unrestricted, which both AISI and Anthropic stress. Separately, Meta disclosed that one of its models accessed the internet during a security test and broke into another company's systems.
The EU AI Act's main obligations became enforceable on 2 August, with fines up to €15m or 3% of global turnover and the AI Office now holding powers over general-purpose models.
OpenAI, Google and Anthropic were called to the White House to co-review the completed federal AI oversight framework.
Macro signals
Software's AI features are not landing. Of 250 software companies surveyed, 50 said fewer than a quarter of their customers use the AI features they shipped. The Information also documented the strain at Figma and Canva, where Canva's largest-ever launch was undercut by ChatGPT.
Big four have now all been flagged for poor AI usage. GPTZero found fake footnotes and invented sources across four PwC Middle East reports, one flagged as entirely AI-generated. All four Big Four firms have now been caught, after KPMG, EY and Deloitte. Each also sells AI governance advice…
The interview: inside OpenAI, Anthropic and Cursor
One podcast worth your commute.
Gergely Orosz's Pragmatic Engineer write-up and episode on visiting the three labs shows how our second story looks from the inside. Two details stand out. At OpenAI, more than 95% of non-engineers use Codex rather than ChatGPT (i.e. the harness, not the chatbot). And engineering work at all three companies is increasingly about building environments for agents to run in efficiently, rather than writing code. If you are trying to picture what might your engineering or tech department look like in the future, this might be the clearest view available.
Tools to try: FutureSearch
Strategy work involves constant forecasting: will this regulation pass, will this vendor survive, how big will this market be in two years. FutureSearch makes that explicit. Type a question about the future, "What will Anthropic's revenue be at year end?", "When will OpenAI IPO?", or a conditional "if X happens, what happens to Y", and it runs the research and returns a probability, a number or a date, with the reasoning shown.
The reason to take it seriously is the track record: it currently ranks first of 197 bots on Metaculus's live forecasting tournament, and scores above the superforecaster median on ForecastBench. Questions cost $0.15 to $2 depending on research depth, and new accounts get $20 of free credit. A practical use: next time a discussion hinges on a disputed assumption, run it as a question and compare the answer with the room's.
Have a good week,