Highlighted Companies
KEY COMPANIESAMD’s data-center segment surged 107% year-over-year to $6.7 billion, driving total Q2 revenue to a record $11.5 billion and topping consensus on both revenue and EPS. Shares still fell nearly 9% in after-hours trading — a reminder that even blowout AI-infrastructure prints are being graded against a much higher bar this earnings season, the same scrutiny investors are now applying to every name, Salesforce included, that leans on an “AI upside” narrative.
Read More →New usage data cited by TechCrunch shows OpenAI closing in on Anthropic’s lead with enterprise customers, months after Anthropic edged ahead among Ramp’s paying business users. The two labs are converging on the same battleground — enterprise seat and API spend — right as both push toward IPOs near or above $1 trillion, meaning market-share swings like this one are becoming a direct input into how each company gets priced.
Read More →AI & Semiconductors
AI & CHIPSTaalas’s approach skips loading model weights from HBM altogether, burning them straight into CMOS to cut the memory bottleneck that dominates inference cost — AMD’s Helios racks, already shipping with HBM4 at 19.6 TB/s per MI455X, would pair GPUs for prompt prefill with Taalas silicon for token generation. Coming on the heels of the MK1, Mext, and FastFlowLM deals, it reads as AMD building an inference-specific moat rather than chasing Nvidia GPU-for-GPU on training.
Read More →Speaking around Nvidia’s push toward a reported $500 billion AI-infrastructure funding plan, CEO Jensen Huang said the company has “largely conceded” China’s AI-chip market to Huawei — a swing from roughly 95% share before export controls to what he now puts at zero. It’s less a near-term revenue story, since China was already a minority of Nvidia’s book, than a structural one: a fully self-sufficient Chinese accelerator ecosystem is now scaling output while competing for the same HBM and advanced-packaging capacity everyone else needs.
Read More →SK Hynix’s lead isn’t just about being first onto Nvidia’s qualification list — its HBM3E yields are now approaching 80%, while Samsung continues working through 12-layer HBM4 stacking issues that keep it around 25–30% of next-gen allocations versus SK Hynix’s 60–70%. Even resolved yield doesn’t fully unlock supply: TSMC’s CoWoS advanced packaging sits between memory production and finished accelerators, so relief depends on clearing both bottlenecks at once, not just one.
Read More →Markets & Tech Stocks
S&P 500 · NASDAQ · MARKETSFactSet’s blended Q2 growth rate of 50.4% marks a second straight quarter above 25% and extends a streak of double-digit earnings growth to seven quarters, with eight of ten sectors posting double-digit gains led by Energy, Communication Services, and Information Technology. Analysts aren’t treating it as a peak either — Q3 and Q4 estimates sit at 27.4% and 25.2% — which raises the bar for every company, SaaS names very much included, reporting into a market that now expects the beat to continue.
Read More →The index closed out the week of August 20 in the red for a second straight week even after jumping 500 points that Friday, a pattern that reads less like a broad risk-off shift than choppy positioning around a market still digesting a red-hot Q2 earnings season and near-record valuations. It’s the kind of week where single-day rallies grab headlines but the net direction is the more honest signal.
Read More →Super Micro added 19% on strong earnings and forward guidance, with Dell up nearly 10% and Micron almost 5% the same week — tangible confirmation that hyperscaler capex commitments are actually converting into shipped server and memory volume, not just showing up as capex line items on an earnings call. It’s the cross-check that makes the $690 billion 2026 capex figure feel real rather than aspirational.
Read More →Supply Chain & Commodities
CHIPS · MATERIALS · FREIGHTJ.P. Morgan now estimates 2026 capex across the five largest US hyperscalers at $697 billion, up $173 billion since the start of the year and nearly double 2025 levels. The financing story is shifting in step — J.P. Morgan’s own framing centers on how AI infrastructure gets funded, through debt, off-balance-sheet vehicles, and partnerships, as much as how much gets spent, because the market’s patience for spend-now-monetize-later is visibly thinner than it was a year ago.
Read More →Samsung’s HBM constraint isn’t wafer capacity, it’s backend packaging throughput — so the company is reportedly considering moving conventional DRAM and NAND backend processing to Vietnam, freeing domestic Korean sites to run as dedicated HBM stacking hubs. It’s a tacit admission that the roughly 50% capacity expansion Samsung announced for 2026 doesn’t matter much until backend throughput, not frontend wafers, stops being the limiting factor.
Read More →TrendForce projects liquid-cooling adoption across AI chip deployments jumping from roughly 33% in 2025 to 53% this year, pushed by chip TDPs now routinely exceeding 1 kW on Nvidia, AMD, and Google’s own silicon. It’s an underappreciated supply-chain constraint sitting alongside the HBM and packaging bottlenecks — data-center buildouts now need to secure cooling infrastructure capacity on the same urgency timeline as the chips themselves.
Read More →