The theoretical ceiling on how fast an eight-GPU Nvidia B200 server can serve one user is 3,047 tokens a second. That number comes from memory bandwidth alone, and real inference engines come nowhere near it. Bandwidth grew, latency did not To compute anything a GPU launches a small unit of work called a kernel, tears it down when it finishes, and launches the next one. Each launch and teardown takes a very short…
Why it matters. If cutting latency were the exclusive territory of purpose-built chips, the ground held by Groq or Cerebras would be safe. If software running on ordinary GPUs can do the same job, what protects that ground is no longer the hardware but the size of the gap that is left.
Testa's monthly disclosure of his long-term set-and-forget account returned to a record. The stated value moved from ¥2,143,066,992 on July 9 to ¥2,196,126,475 on August 9. The account screenshots are public, but this is not an audited fund report, so “record” should be read as a claim within his disclosed series. The arithmetic does check: the eight rows sum exactly to his total. Two rows made the record The accou…
Why it matters. The source of the gain matters more than the record total. Two buckets supplied 88% of the increase, while several index labels can overlap in the same US mega-caps, leaving effective risk more concentrated than the row count suggests.
An AI accelerator can sit idle when data does not arrive fast enough, even if its compute chip is fast. Samsung's zHBM and the HBF project led by SK hynix and Sandisk both attack that delay, but they are not direct substitutes. One shortens the path to fast memory; the other adds a much larger nearby storage layer. zHBM was already public in February Samsung first described zHBM at Semicon Korea on February 11, 202…
Why it matters. The investment difference lies in product maturity, not peak multipliers. HBF has a specification and sample dates but limited adoption; zHBM has an integration advantage but no named customer or production date.
Samsung is reported to have lifted its HBM4 yield to 80%. When mass production began this February the figure was still under 60%, so this is a six-month change. The same report says the next generation, HBM4E, cleared a 70% yield in reliability testing, and that 12-layer samples went out to customers in May. A target set for year-end arrived four months early. Why yield turns into money so directly HBM is DRAM s…
Why it matters. Yield is a number a company can move inside its own fab, but market share is set by where customers place their orders. The DRAM-level HBM share Samsung has put on the table for year-end will be decided by Nvidia's volume allocation, not by the yield line.
Unrealized losses on domestic bonds held by Japan's four largest life insurers reached a record 15.13 trillion yen, about $96 billion. That is 7% more than the previous quarter, and the seventh straight quarterly increase. The figure Bloomberg reported first covers Nippon Life, Dai-ichi Life, Sumitomo Life and Meiji Yasuda together, and all four saw their paper losses grow. Nippon Life alone accounts for 6.28 trill…
Why it matters. The record number makes Japanese life insurers look fragile, but the yardstick for measuring their solvency changed this year to one that marks both assets and liabilities. Depending on which yardstick you pick up, the same rise in rates reads as a crisis or as an improvement.
US federal debt does not disappear after one budget cut. Tax and spending choices can recreate the deficit every year. Steve Hanke argues that breaking this cycle requires a constitutional debt brake, proposed through the state-led route in Article V. Article V has two gates The Constitution offers two ways to propose an amendment. Two-thirds of each chamber of Congress can act, or two-thirds of state legislatures…
Why it matters. A debt brake could lower long-run fiscal risk in Treasuries, but markets will treat it as a slogan if the Article V path and the rule's design remain vague. State applications and the treatment of exceptions are the tests that connect it to rates.
The US Strategic Petroleum Reserve has physically declined. EIA data show inventory falling from 415.4 million barrels at the end of February 2026 to 311.4 million on July 17. That is a drop of about 104 million barrels, or 25%, in five months. This is a crude loan, not an outright sale Doomberg's distinction lies in the contract structure rather than today's custody of the oil. In March, the Energy Department ann…
Why it matters. An SPR exchange moves today's oil shock into a future repayment burden. Oil and refining investors need to watch both the current inventory draw and next year's crude demand for returns.
A four-year high in lumber would normally suggest that more homes are being sold. This time the opposite is true. US new single-family home sales were lower than a year earlier in June, while framing-lumber prices rose. Housing demand and lumber prices split apart The industry framing-lumber composite reached $535 per thousand board feet in July. That was the highest since September 2022 and 23% above a year earlie…
Why it matters. Builder costs can rise when lumber supply falls faster than home sales. Even lower tariffs may not remove price pressure if closed mills do not return. That is why demand data alone cannot explain margins in builders or lumber producers.
Serenity models a $427 million midpoint for Sivers' annual revenue. The thesis is that new capacity can be absorbed while laser supply remains tight. The arithmetic is attractive, but company guidance and an investor-built capacity model are different things. Where $427 million starts The model assumes Sivers receives 10% of WIN Semiconductors' wafer capacity. It then applies a 65% yield and a $50 to $75 average se…
Why it matters. The laser shortage can be real while Sivers' revenue still depends on wafer allocation, yield, and price. Reading $427 million like guidance before the company supplies those inputs confuses possible capacity with booked revenue. New shares must also be reflected in the denominator of any valuation multiple.
Three macro numbers carried by the Kobeissi Letter last week did not move in one direction. US companies earned near-record profits, China's central bank bought more gold, and information-sector layoffs rose, but by less than the post claimed. Risk appetite, demand for insurance, and labor stress coexisted. Check the profit denominator Kobeissi wrote that US pretax corporate profits reached 14% of GDP. The exact ra…
Why it matters. Corporate profits, central-bank gold, and sector layoffs move different markets. Together they show a mixed regime: risky-asset cash flows remain strong while institutions and firms add insurance. Correcting the official layoff rate to 2.0% materially weakens the recession narrative.