There is a version of this week that reads as a quiet drawdown: SPX down 0.8%, BTC down 3.1%, gold down 2.0%. Nothing dramatic. But October crude gained 10.1% to $100.62 and the ten-year real yield rose 17 basis points to 2.60%. And so the quiet week turns out to be two squeezes running at once, in the two inputs that price everything else. Fuel got more expensive and so did money. Most of the assets on this page were simply standing in the way.

Oil's move was not precisely because of more demand. Saudi Arabia confirmed a precautionary shutdown of the East-West pipeline on Sep 11, the route Aramco's own Aug 4 results credited with keeping exports flowing through the Hormuz disruption. The market answered. Friday's preliminary settlements put October WTI at $100.05 against $80.11 for April 2027. That is a $19.94 premium for a barrel now versus a barrel in the spring.

Meanwhile, the rates move was already underway before Friday's data. Of the ten-year nominal yield's 18-basis-point rise, 17 arrived before CPI printed. That timing means the bond market was not reacting to the inflation number, it was positioned ahead of it. Real yields did most of the work, rising 17 of those basis points, and a higher real yield is a direct tax on anything valued off distant cash flows. Gold, which pays nothing at all, fell 2.0% in the same week funds put $18B into it.

Underneath the indices, the market stopped moving as one thing. SPX fell 0.8% while equal-weight RSP fell 1.9% and small-cap IWM fell 2.4%, so the further you got from the largest companies, the worse it was. SOXX, however, gained 1.4%, but even inside chips the direction split, with NVDA down 5.2% while QCOM rose 7.8% on an AMZN supply agreement. Crypto did the same thing. Thirteen of the 26 tokens covered here rose and thirteen fell, with computing-linked names leading and several lending and trading protocols lagging.

So this was not a risk-off week. It was a week where two costs went up and the market sorted its holdings by who can pay them. The names that gained were mostly selling something scarce or collecting a fee that grows with activity. The names that fell were mostly financing something, discounting something distant, or burning fuel. The Fed decides on Sep 16, and Sep 16 retail sales arrive the same day, covering August spending that predates the confidence drop households reported this month.

Key Takeaways
- Oil and real yields rose together, and the market sorted every other holding by who pays those two bills. October WTI gained 10.1% to $100.62 while the ten-year real yield rose 17 basis points to 2.60%. SPX fell 0.8%, BTC 3.1% and gold 2.0% standing in the way of both.
- Seventeen of the ten-year's 18-basis-point rise happened before Friday's CPI; the bond market was positioned ahead of the data rather than reacting to it. Polymarket's Sep 16 hike contract still moved from 59.5% to 79.5% in the two minutes around the release.
- Crypto split down the middle, 13 of 26 tokens up and 13 down, and the fund channel split with it: US spot BTC funds lost $462.7M across four sessions while ETH funds gained $196.9M. Compute-linked tokens led, fee businesses with falling revenue lagged.
- The market stopped paying for AI adoption metrics. TTAN fell 37.8% and BRZE 25.1% after guiding to roughly flat sequential revenue, and ADBE fell 5.4% while raising guidance on AI recurring revenue it declined to size.
- US gas is the anti-oil trade. Its cushion narrowed to 148 Bcf above the five-year average from 160 Bcf a week earlier, while oil buyers paid a $19.94 premium for October delivery over April 2027.
Macro: Real Rates Lead the Move
August inflation handed the Fed a genuinely split picture. Headline CPI rose 0.4% on the month against July's 0.1%, with gasoline's 3.9% rise supplying more than a third of the increase. Core accelerated to 0.3% monthly even as the annual core rate eased to 2.4%, and PPI excluding food, energy and trade services slowed from 0.4% to 0.3%. The pipeline is cooling while the pump is not. That is a familiar problem with an unfamiliar cause, because this oil move is a supply interruption rather than demand. Claudia Sahm, the economist behind the Sahm rule—which flags the onset of a recession when the three-month average unemployment rate climbs half a percentage point above its low of the prior year—had already said on Sep 10 that she favored a hike over a hold.

The market's answer was immediate. Polymarket's contract for a 25-basis-point hike on Sep 16 moved from 59.5% just before CPI to 79.5% within two minutes of it. A twenty-point repricing in 120 seconds is the sort of thing that only happens when positioning was one-sided going in.

The bond market had already committed. The ten-year real yield rose 17 basis points to 2.60% and the nominal ten-year gained 18 basis points to 4.96%, with 17 of those nominal basis points in place before Friday. On Friday itself the two-year climbed seven basis points against the ten-year's one, with the curve doing the flattening a hike expectation produces. Real yields, not inflation expectations, carried the move, which is what makes it expensive for everything priced off long-dated cash flows. Scarce cash was not the cause. Weekly-average bank reserves rose $96.8B to $2,991.3B for the week ended Sep 9 while the Treasury's account at the Fed fell $84.6B to $883.3B, which is the Treasury paying its bills and moving that cash into commercial bank accounts. Banks held more settlement cash and bondholders still took the mark, because reserves answer to the payment calendar and bond prices answer to the real yield.

Households are already behaving as though the squeeze is real. Michigan's year-ahead inflation expectations jumped from 4.0% to 4.6% in preliminary September results, and the broader expectations index fell 11.1% to 45.8, with the survey linking fuel prices and trade tensions to household budget worries. The labor market disagrees. Initial claims edged down to 206K, and longer hours lifted August real weekly earnings 0.2% despite a 0.1% fall in real hourly pay. People are working more hours to stand still, which is a description of a cost shock rather than a downturn.

The rest of the world is not waiting for the Fed. The ECB announced a 25-basis-point rise across all three policy rates on Sep 10, effective Sep 16, taking the deposit rate to 2.50% and citing inflation pressure from the Middle East conflict, which puts a second central bank tightening into the same oil shock on the same day. Growth elsewhere was uneven. UK GDP grew 0.4% in July while consumer-facing services shrank 0.4%, and China's August CPI rose 0.8% year over year against a 3.8% increase in PPI. Chinese factory-gate prices running almost five times its consumer prices means the world's largest goods exporter is no longer exporting disinflation, and that is a support US goods prices have been leaning on.

Crypto: Selective Gains, Uneven Business Growth
Crypto did not have a direction this week, it had a sorting. Thirteen of the 26 tokens covered here rose and thirteen fell, and the dividing line was not size or beta. NOCK gained 78.2% and VVV 32.2% at one end; PUMP lost 16.1% and CHIP 14.9% at the other. BTC fell 3.1% while ETH rose 3.4%, a 6.5-point gap between the two assets that usually move together, so even the top of the market split.

The clearest split ran through the fund channel. US spot BTC funds recorded four consecutive negative sessions totaling $462.7M of outflows, reversing the prior week's $986.7M of creations, while US spot ETH funds took in $216.4M on Friday alone and finished the week with $196.9M of net creations. Same wrapper, same access, opposite direction. When one route produces opposite flows for the two largest assets in the same week, the driver is allocation preference.

What actually worked was compute and access, proxying the AI narrative. Tokens attached to selling or renting compute led: NOCK on reproducible inference, VVV on API credit capacity that requires locking the token, NEAR up 15.7% on an optional staking-for-compute program, TAO up 5.1% on a subnet that billed $964K to 1,187 paying renters. What lagged was the fee businesses whose income fell, with PUMP's bonding-curve revenue contracting 38.6%, HYPE's holder revenue down 12.7% to $11.57M, and LIT's down 37.2%. Price and business came apart in both directions. UNI gave back 1.9% while its protocol revenue rose 46.5% to $5.13M.

The week's product news landed almost entirely in one contested place: stock-linked trading. Pump launched Custom Pairs on Sep 9, PONS answered with ten more stock pairs on Sep 11, Pendle opened markets for fees earned by selected PONS-launched tokens, and Token Terminal counted $3.8B of bStocks volume on BNB Chain over thirty days. Just an example of how four protocols are now chasing the same pool of traders who want equity exposure onchain. Expect more to come.


The compute trade is where crypto and the AI capital cycle finally touch. Lium billed $964K to 1,187 renters on Bittensor, Venice takes its DIEM capacity target to 40,000 on Sep 14 with each staked unit funding $1 of daily API credits, and NEAR now sells $5 of monthly hosting credit for 500 staked tokens. Nockchain added the research layer with reproducible inference across devices, at 1.32 times baseline throughput on one model and 0.52 on another.

BTC: BTC ended the week at $77,282, down 3.1%, and the fund channel reversed on it. US spot BTC funds took $462.7M out across four consecutive negative sessions, undoing the prior week's $986.7M of creations. The shape of the outflow says more than the total: Thursday alone was $282.7M and Friday only $13.2M, so the selling was front-loaded and fading by the close. What makes this week genuinely informative is the contrast with ETH, which took in money through the identical channel. BTC sits 38% under $124,740, and the same fund channel that carried it there is now the one selling. A positive creation session before the next Thursday would mark this as a holiday-week air pocket; a second consecutive negative week with ETH still gathering makes it a rotation, and rotations inside a single asset class run further than people expect.


ETH: ETH closed at $2,538, up 3.4%, beating BTC by 6.5 points, and one session did the work. US spot ETH funds took in $216.4M on Friday after the prior three sessions bled $19.5M, flipping the week to $196.9M of net creations. Friday alone was larger than the entire week's net, which is the detail worth holding: a single print of that size is either one allocator rebalancing or the start of a trend. ETH now trades 46% below its $4,691 high, doing the recovering while BTC stalls. A second week of creations makes this institutional. Friday's inflow reversing makes it one desk, and the week's outperformance becomes positioning noise wearing a flows story.


BNB: BNB managed just 0.7% to $724.87, the weakest gain among the majors. Token Terminal recorded $3.8B of bStocks DEX volume over thirty days, roughly $3.1B of it on PancakeSwap. Tokenized stocks are cheap to trade and get churned, which is exactly the combination that produces enormous notional with trivial fee spend. The product is real and active; the token capture is unproven. Watch chain fee revenue, not notional. Rising fees mean BNB participates. Flat fees beside climbing volume prove the turnover free-rides on a chain that discounted itself. Thirty days of bStocks volume at $3.8B has not moved the fee line. Volume without fees means someone else's business running on your chain.

SOL: Solana held at $102.01, up 0.3%, waiting on a change that has not shipped. The relevant mechanic is rent: a refundable SOL balance an application must lock to keep account data live on the network. Solana cut that requirement 9% on Sep 3 and targeted mid-September for a second stage taking the total reduction to 27%. Lower rent means a developer locks 27% less SOL to open the same account once both stages land, which removes a quarter of that structural demand while making the chain cheaper to build on. SOL is 59% under $247.60, and the change that would matter has not shipped. Until it does the price waits on a governance calendar.

ZEC: Zcash gained 12.9% to $1,174.88 and is now within 5.6% of its 52-week high, the tightest of any token in this edition. Privacy found a bid while the majors stalled. The work behind it is unglamorous and the right kind: a Sep 10 update reported Zakura network-health improvements, faster shielded wallets, and adoption by mining pools. Shielded payments have always been Zcash's product and always been too slow to use (wallet speed is the adoption constraint, not cryptography). The Foundation's advisory NU7 poll closes Sep 14 at 19:00 UTC.

HYPE: HYPE fell 4.8% to $80.79 while its fee-funded holder revenue fell 12.7% to $11.57M, and for once the token and the business moved together. That is worth noting because most of this section is price detaching from fundamentals. The token sits 8.1% below its $87.95 high and 283% above its low; the market still prices it as the best-performing fee asset in crypto.

UNI: UNI fell 1.9% to $6.08 after a 41% rally the week before; the pullback had nothing to do with the business as protocol revenue rose 46.5% to $5.13M in the same week. Uniswap runs the largest onchain exchange and converts collected protocol fees into UNI it burns, with the fee switch live. So a stronger fee week feeds the burn directly, and the token still trades 40% below its $10.13high. That gap between rising fees and a retracing price is the Robinhood Chain trade unwinding faster than the fundamentals did.

NEAR: NEAR rose 15.7% to $2.51, the second-best major this week. Venice named NEAR AI Cloud and Phala Network as providers for its hardware-protected AI service and NEAR launched an optional staking-for-compute program: lock 500 NEAR and receive $5 of monthly agent-hosting credits. The credit is funded by staking yield and token price, so a holder is converting variable yield into a fixed credit while keeping full price exposure on the principal. If the token falls, the staker owns the drawdown and still gets $5.

TAO: Bittensor rose 5.1% to $236.11. Lium, its GPU-rental subnet, reported $964K billed to 1,187 renters in its monthly review. Lium's rules separate customer-funded rental income from newly issued subnet tokens that reward providers, while TAO holders sit at the network level and have no claim on Lium's rental receipts at all. Subnets' customer revenue is not TAO demand, but it’s a good sign to see an increasingly larger number of enterprises relying on ecosystem subnets and open-source AI. $964K billed to 1,187 renters buys the only paid demand in this section that arrives with customer names attached.

AAVE: AAVE fell 4.7% to $124.48 in the week it announced more than $900M of V4 deposits. Deposits are supply, and supply earns nothing in lending until somebody borrows it. But the supply-side has to show before borrower demand can come in. For now, most of TVL still sits on V3. Aave is the largest onchain money market, and its revenue comes from the spread between what borrowers pay and what depositors receive. The puzzle pieces are slowly coming together and it seems reasonable to expect an increase in the amount of fee revenue allocated to buybacks as we approach the end of the year. The buyback allocation is the year-end question. $900M of V4 deposits forms the supply that has to earn it first.


MORPHO: MORPHO fell 8.0% to $2.33, the worst week among the lending tokens. Morpho's August recap reported more than $14B of deposits and over $400M sitting in Robinhood Earn as it continues to expand its roster of distribution partners. MORPHO remains 20% below its high, the shallowest drawdown among the lenders here.

ENA: ENA fell 12.9% to $0.1424, the steepest decline among the larger tokens as the news of its fee switch and neobank focus were digested. This week, right after the release and emphasis on Ethena Pay, Ethena announced USDe on TRON on Sep 11 with bridging for USDe and sUSDe opening the same day through Stargate, plus integrations into other TRON products over subsequent weeks. Tron is the largest chain for stables, namely USDT, and Ethena is now attempting to gain a fair share of that market. Distribution is genuinely the constraint for a dollar-neutral position like Ethena: the product is only useful where people already hold balances. ENA is 81% below its $0.76 high, the deepest drawdown in this section, though supply is the number that decides this one. The fee switch takes 5% of gross revenue once USDe circulation reaches $7.5B and nothing at all below it. The test is the trailing 14-day average rather than a spot print that can cross for a day. That average sits at $4.4B against $4.60B spot, so supply has to grow 70.6% to turn it on, and it grew 16.3% over the past month. That is what the TRON listing is for.


PUMP: PUMP lost 16.1% to $0.003598 while bonding-curve revenue contracted 38.6%, and the buyback kept running straight through it. The program allocates half of specified revenue to buying and burning tokens, on a one-year term from April 2026, which means the clock runs out in April 2027 whatever happens to the price. That is the structural point: this is a finite program that suffers when it’s funded by a shrinking revenue line which may find no permanent floor. Custom Pairs, launched Sep 9, lets new tokens trade against tokenized stocks, and PONS answered with more stock pairs two days later; the fee pool Pump depends on is now contested as memecoins thrive elsewhere on Robinhood Chain. Bonding-curve revenue stabilizing while custom-pair fees ramp is what funds purchases to the April 2027 horizon; revenue falling continually makes the buyback a countdown.


ARB: ARB gained 7.5% to $0.1416. Arbitrum is now the third biggest chain by revenue, all of which goes back to the token-controlled treasury. Robinhood Chain stands at the second biggest, capturing ~90% profit margins on all transactions flowing through their dedicated Arbitrum chain. The Foundation disclosed $6.19M of first-half DAO income from transaction fees, Timeboost, expansion-program licensing, and treasury income. Four sources behind $6.19M, only one of which is ordinary chain activity, which makes the treasury more diversified than the usual rollup economics. Timeboost in particular sells transaction ordering, a revenue line most chains give away. ARB governs a treasury and holders direct capital (they do not receive it unless governance votes to distribute or do buybacks). Second-half DAO income growing beyond $6.19M with a distribution mechanism attached would turn governance into cash flow; income growing while the treasury only accumulates leaves ARB a voting right on someone else's balance sheet.

LIT: Lighter slipped 3.6% to $4.47 and sits 7.4% under its 52-week high. The venue's fee base meaningfully stepped up on Aug 19, from about $78,000 a day to about $600,000, and has held there since. Fees ran $5.44M over the 30 days to Sep 11 against $2.37M in the 30 days before, up 129% month over month. This same week reads the other way. Fees fell 12.0% and holder revenue fell 38.5% against the prior week. Holders captured $3.06M of those 30-day fees, 56% of the total. Lighter is the perps venue on Robinhood Chain, which is where the order-flow case rests.


PONS: PONS fell 13.2% to $0.6008 despite recording $10.46M of protocol revenue and $7.93M of burn-valued holder revenue, which would be an extraordinary week for almost any token here. Meanwhile the competitive picture tightened: PONS announced ten more stock pairs on Sep 11, two days after Pump's Custom Pairs, so both platforms are now chasing the same stock-linked trading fees, the former on Robinhood and the latter on Solana. Revenue near $10M through a contested quarter would prove a durable fee base, but falling fees as pairs multiply means the whole category is splitting one pool of traders.

AERO: Aerodrome gained 15.9% to $0.5806 on $9.18M of voter income, and one day supplied almost all of it. Fees on Sep 9 came to $8.18M. Uniswap earned $52.75M over the same seven days with no spike in it, its fees down 16.0% week over week while Aerodrome's rose 542%. Aerodrome still grew fees 225% month over month against Uniswap's 83.5%, and UNI outran AERO on price anyway, 68.2% to 38.5%. What holders are paying for is the merger with Velodrome into a single Aero spanning Ethereum and Circle's Arc, with AERO holders taking 94.5% of the unified supply. Arc's mainnet goes live on Sep 16.

PENDLE: Pendle rose 4.1% to $2.01 on revenue up 194.5% against the prior week. Over the thirty days to Sep 11 fees came to $656,523 against $699,961 in the thirty before, down 6.2%. The base is small because the protocol shrank. TVL is $1.24B against $12.05B a year ago. Boros, the funding-rate product the team points at for growth, billed $34,590 across those same thirty days and fell 11.7% month over month, which leaves it near 5% of Pendle's fees. Eighty percent of revenue has gone to stakers since September 2025, and vePENDLE became liquid sPENDLE on Jan 20 with a fourteen-day exit, a split that has paid $14.91M over the trailing year. The new markets are Pons assets on Robinhood Chain, SHROOM and microDUCK first, with the opening maturities set for Sep 24. Every Pons token pays a creator tax on trading plus a holder fee share, so Pendle is splitting a memecoin's trading fees into a fixed leg and a floating one. It essentially prices two weeks of launchpad volume as though it were a yield curve.

AI: Artificial Inu gained 29.6% to $0.2858. The meme pairs its token with tokenized NVDA and runs a fee-funded vault on Robinhood Chain, which is going through its own meme season. Nvidia itself fell 5.2% this week while this token rose 29.6%.

DRV: Derive slipped 1.2% to $0.1489 and bought its own token anyway, for the 83rd consecutive week. The Sep 8 purchase took 209,724 DRV at an average $0.16, lifting cumulative purchases to 27,445,701 tokens. The venue allocates 35% of protocol fees, so the size of each purchase tracks what the options and derivatives business actually earns, not the notional of contracts traded. DRV sits 15% under its high, the tightest chart among the fee tokens after LIT. The Sep 8 purchase averaged $0.16 against a $0.1489 close, 7.5% over the market it was supporting.

CARDS: Collector Crypt's token fell 12.2% to $0.1260 in another week its business got measurably better; that divergence is actually the normal state rather than the news. Across the last twenty-five weeks price and tracked revenue moved the same direction ten times, and the correlation between their weekly changes is −0.07. This token has never followed its own business. Net revenue rose 37.9% to $2.23M after card buybacks (customers pay to open packs and can sell cards back, so gross pack sales include money that flows straight out again). Weekly revenue has run between $1.31M and $5.24M over six months against a median of $2.40M. The trailing year came to $85.0M against a $120.5M market value, so the market pays about 1.4 times revenue for a business it will not reprice week to week.

CHIP: CHIP fell 14.9% to $0.04728 while the product it governs kept growing. USD.AI reported $464M of sUSDai on Sep 9 and quotes a 7.02% net annual yield on its GPU-backed lending product. That 7.02% belongs to the interest- bearing dollar, not to CHIP (a holder of the token receives no part of it; CHIP governs risk parameters and fee allocation, which means it sets the terms of the credit rather than collecting the interest). The chain runs in three stages: customers supply dollars, loans fund GPU borrowers, and governance prices the risk.

HNT: HNT fell 14.2% to $0.5109, which is the give-back after last week’s run on a short squeeze. The upside move was Aug 30 and 31, when the token ran from $0.2526 to $0.7663 in two sessions, up 203%, after Celina, Texas converted its municipal Wi-Fi to carrier coverage on the network. Reported volume went from about $4M on Aug 28 to $230M on Aug 31, and the largest single-day short squeeze HNT has recorded this year did the rest. That repricing added roughly $97M of market value for a deployment Helium puts at 100 GB a day, which annualizes to 36.5 TB. The token has since given back 37% from the Aug 31 peak. Three weeks before that run, the supply side turned. Data Credits pay for all service on the network, and burning HNT is the only way to do that. That burn ran ahead of issuance until Aug 6, retiring about 22,000 HNT a day. On Aug 7 the HIP-149 supplement started minting. The float has grown 7.6M HNT since, close to 200,000 a day. The supplement runs flat for 12 months, then tapers over 24, about 141M HNT in all. It lifts the effective ceiling from about 206M to about 347M. Set Celina against that. At Helium's rate, 100 GB a day bills out near $45,000 of Data Credits. Issuance now runs about $102,000 a day at this price. The deployment that doubled the token does not pay for half a day of it.

NOCK: Nockchain rose 78.2% to $0.03789, the largest gain in this edition. The paper behind the story arrived after the move. The decisive session was Sep 6, up 51.3% on volume that tripled to $3.9M. What came before it was a compute claim. Nockchain switched its miners to AI proof-of-work in mid-August. On Sep 4 it put the network's peak work rate at 19.6 ExaMAC/s, about 19,600 H100 equivalents, and 170 times the first reading three weeks earlier. On Sep 7 it valued the hardware behind each block at more than $435M. The token opened the week at a $51.5M market cap and finished near $83M. Lockstep, the bit-exact inference paper (verifiable inference across devices), landed Sep 10, three days after the repricing. The check on all of it is where NOCK trades. An Aerodrome pool on Base carries about 65% of reported volume, and Kraken carries $148K.

VVV: VVV gained 32.2% to $24.09 and now sits 6.9% from its 52-week high. Minting DIEM requires locking staked VVV, and each staked DIEM supplies $1 of daily API credits; capacity and token lockup move together by construction. Venice's next scheduled increase takes the DIEM target to 40,000 on Sep 14, which creates room for $40,000 of daily credits if fully minted. The company also announced a discretionary burn valued at $391K on Sep 8; discretionary is the key in that sentence, because a burn management chooses is not a burn the mechanism guarantees. Minted DIEM approaching 40,000 after Sep 14 is paid demand. The target rising while minting lags means capacity is being built ahead of customers.

JUP: JUP rose 12.0% to $0.2431 on Clans, a feature that treats prediction markets as a social product rather than a financial one. Launched Sep 8, it lets groups share calls and compete together, so a user comes back to see how their group is doing rather than only their own position. That is deliberate retention design, and retention is the scarce thing in prediction markets, where most volume arrives around events and leaves with them. By Sep 9 there were 46 clans and Jupiter still called it a trial run. Prediction markets are also the newest counter in a business that keeps adding them. Jupiter is a router first, with perps, lending, a launchpad, and a mobile app on top of that routing, and a stocks page went into preview on Sep 10. The older counters carry the economics. Fees were $4.25M for the week, down 2.8%, but the 30-day book is $19.9M, up 71.0% on the prior 30 days and annualizing near $243M. Revenue, what Jupiter keeps after paying liquidity, was $1.84M for the week and $7.07M over 30 days, up 86.7% month on month, which annualizes at $86M or 35% of fees. Half of onchain revenue buys JUP through the Litterbox Trust, which burns it, about 134M so far and equal to 4% of the float. That is $43M a year against a $781M market cap, so the token trades at 9 times annualized revenue and 18 times the half that reaches holders.

Equities: Same AI Demand, Different Rewards
The US market narrowed again this week. SPX, the S&P 500 index, fell 0.8%, equal-weight fund RSP fell 1.9%, and small-cap fund IWM fell 2.4%. Stretch that to four weeks from Aug 14 and the gap widens: SPX down 1.7% against RSP's 3.5%, twice the damage once you stop letting the largest companies carry the average. This is a market where the index is healthy and the median company is not.



SOXX gained 1.4% and immediately disproved the idea that AI trades as a bloc. NVDA fell 5.2% while QCOM rose 7.8%, a 13.1-point spread in one week, after QCOM and AMZN announced a multigeneration agreement covering custom inference silicon and optical connectivity, and SemiAnalysis put Google's Ironwood at $0.181 per million tokens against NVDA's B200 at $0.222. The AI budget is not shrinking. It is being redistributed, and this week the market started naming the recipients.




For the companies buying that capacity, the bill arrived in public. ORCL grew cloud-infrastructure revenue 121% and still outspent operating cash flow by $5.4B, covering the gap with $19.9B of net equity issuance on top of $11.4B of customer prepayments already inside the operating line. CRWV finances GPUs at SOFR plus 4.5 points against two non-investment-grade customers. IREN carries $4B of contracted revenue against $70.5M recognized. Growth is not the constraint in this group; funding it without diluting or levering the owner is.



The software results drew the sharpest line of the week. TTAN fell 37.8% and BRZE 25.1%, both after reporting solid quarters and then guiding to roughly flat sequential revenue. ADBE fell 5.4% while raising full-year guidance, having grown AI-first recurring revenue more than 150% without disclosing the base it grew from. The market is no longer paying for AI adoption metrics. It is paying for the revenue line those metrics were supposed to produce, and three companies discovered that in the same week.



Meanwhile, the consumer platforms are trying to charge for attention they already own. META launched Muse on Sep 8, free at entry with paid tiers, running on a dedicated virtual machine whose data META says stays out of its advertising systems, which closes the easy monetization route. AAPL put iPhone 18 Pro at $1,199 and Pro Max at $1,299 with availability from Sep 18 and Siri's English beta landing Sep 14. One company is selling a subscription into an audience it already has; the other is selling hardware to the same base every eighteen months. Both depend on conversion rates neither has disclosed.



SPX: The S&P 500 lost 0.8% to 7,656.98, and Friday's 0.9% rebound was not enough to save the week. The number that matters is not the index, it is the spread underneath it. Over the four weeks from Aug 14 the S&P fell 1.7% while the equal-weight version fell 3.5%, twice as much. Cap weighting is doing the work: a handful of the largest companies are holding up a benchmark that most of its own constituents are dragging down. That is a market narrowing, not a market falling, and the two require different positioning. The index sits just 1.8% below its $7,799 high and 21% above its low, so the headline still looks healthy. The spread closing, with equal weight catching up, confirms the weakness was rotation rather than deterioration. The spread widening past four points says the rally has shrunk to a handful of names, which is when index exposure stops being diversified exposure. Buying SPX here is buying concentration.

NDX: The Nasdaq-100 slipped 0.6% to 29,368.44 as its constituents also refused to trade as one thing. Nvidia fell 5.2% in the same week Apple gained 3.8% and Meta 5.1%, an 8.3-point spread inside a bloc the market prices as a single AI bet. The index holds three different businesses wearing one label: companies that sell computing, companies that buy it, and companies that distribute products using it. Apple's Sep 9 hardware launch and Meta's agent both monetize the last of those, and neither depends on accelerator demand. The index is 4.2% below its $30,661 high and 28% above its low. Dispersion staying this wide is the signal to stop trading the index and start trading its parts; the names reconverging means the AI trade is still one factor and position size should reflect that. Right now it is three trades in one ticker.

SOXX: Semiconductors gained 1.4% to $527.07, and the bounce is smaller than it looks. The ETF is still 4.2% below its Aug 14 close and 19.5% below its $655 high, while sitting 106% above its 52-week low. That is a violent, unfinished round trip, and it tells you the AI capital budget is being redistributed rather than cut. Qualcomm's Amazon deal is the redistribution in action: custom inference silicon and optical connectivity taking share of the same spend. Memory and data movement are the other claims on it, and Marvell already printed 37% revenue growth. The near test is dated: Micron reports Sep 30, the first read on whether memory pricing is capturing the budget or merely riding it. Reclaiming the $550 Aug 14 close puts the sector back in an uptrend; failing there with Micron disappointing confirms this bounce as a lower high inside the decline. Chips are the highest-beta way to be long AI and the fastest way to be wrong about who collects.

RSP: Equal-weight large caps fell 1.9% to $214.87, more than twice the S&P's 0.8%, and the four-week gap speaks by itself, down 3.5% from Aug 14 against the cap-weighted index's 1.7%. Same companies, different weights, double the damage. RSP gives every large constituent a similar weight, so it cannot lean on a few mega caps to carry a bad week, and its sector mix differs enough that identical earnings news lands differently in the two funds. Friday's 0.8% rebound did not change the direction. The fund is 3.6% below its high and 18% above its low. Watch the ratio: RSP outperforming SPX for two consecutive weeks means breadth is repairing and the average stock is participating again. RSP lagging by another two points says the index is a concentration vehicle wearing a diversification label.

IWM: Small caps dropped 2.4% to $288.89 and are 5.3% below their Aug 14 close, the worst four-week showing of the three US equity wrappers on this page. There is a lazy interpretation that can be made about small caps and rates, and it is worth resisting. The Russell 2000 is not a floating-rate-debt basket or a pure domestic-consumer bet: its members carry different debt terms and refinancing calendars, so a rate move reaches a company with fixed funding on a completely different schedule from one refinancing next quarter. The index return blends those effects with earnings revisions and sector mix, and pulling them apart is the work. IWM is 26% above its low and 5.3% below its high. Small caps outperforming on a week when the ten-year falls confirms the rate channel; continuing to lag while yields fall means the problem is earnings, not financing.

Mag7: The equal-weight Mag7 basket gained 0.8%, and the average hid everything interesting. Meta rose 5.1% and Apple 3.8% while Nvidia fell 5.2%, a 10.3-point range. This basket weights Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, and Tesla equally at the start; it is not a quoted index, and that construction is what surfaces the dispersion a cap-weighted version would bury. The split is not random. Meta launched a consumer agent, and Amazon widened its chip suppliers, so within one week the group contained a seller of computing losing ground, a buyer of computing gaining bargaining power, and a distributor monetizing neither. Dispersion compressing back under four points means the AI trade is one factor again; staying above ten points means owning the seven as a bloc is a decision to hold both sides of the same argument.




NVDA: Nvidia gave back 5.2% to $218.29 and erased almost all of the prior week's 5.9% gain. Qualcomm and Amazon announced custom inference silicon and optical connectivity, and SemiAnalysis put Google's Ironwood at $0.181 per million tokens against $0.222 for Nvidia's B200 at a matched operating point. That is a 19% modeled cost gap on one workload, and it is the first time the comparison has been public and specific. The stock is 7.4% off its $235.74 high and 32% above its low; the market is pricing competition, not decline. Micron's Sep 30 print is the next industry read on whether the AI budget is still growing or merely being redistributed. A top-four cloud buyer disclosing a second-source commitment at scale would confirm the redistribution; a hyperscaler raising its capital plan without naming an alternative supplier kills it. Owning Nvidia is a bet on the size of the budget. Owning it against Qualcomm or Broadcom is a bet on who collects it.

AAPL: Apple rose 3.8% to $332.27 and is now the healthiest chart among the mega caps, 2.3% from its $340 high while Alphabet sits 16% below its own. The catalyst is dated and priced: iPhone 18 Pro at $1,199 and Pro Max at $1,299, with availability from Sep 18, Siri's English beta arriving with iOS 27 on Sep 14, and preorders opening Sep 12 after Friday's close. Importantly, Apple sells devices and then rents services on top of them, so the upgrade cycle matters twice. Premium pricing plus carrier credits can move units without moving realized revenue per phone, which is the number that actually reaches earnings.

GOOG: Alphabet finished the week unchanged at $335.45, up 0.04%, and it is the most interesting flat print on this edition. SemiAnalysis put Ironwood at $0.181 per million tokens against Nvidia's B200 at $0.222 in a matched 100-tokens-per-second comparison, which makes Alphabet the only company here sitting on both sides of the AI cost equation: it buys compute and it designs the alternative. The stock trades 16% below its $399 high, the widest mega-cap discount on the page, and 41% above its low. Cloud margin expanding in the next quarterly print converts the silicon advantage into money; margin flat while capital spending rises means the TPU is a cost center with good benchmarks.

MSFT: Microsoft slipped 0.8% to $495.63 and finished the week exactly where it sat on Aug 14, up 0.05% over four weeks while the equal-weight market fell 3.5%. Flat is outperformance in this regime. The operating news is that CFO Amy Hood said on Sep 9 that some dock-to-live times had fallen by more than half over a year. Dock-to-live is the lag between a server arriving and earning Azure revenue, so halving it pulls the revenue on already-purchased hardware forward without buying a single extra chip. That is the most valuable kind of efficiency in a capital-intensive build, and Microsoft had already flagged comparable gains in July, so this is a trend confirming rather than a surprise. Commissioning is only the first gate. How much of that capacity runs, and how much of it is paid for, decide what it earns.

AMZN: Amazon fell 0.7% to $256.78 and spent the week rearranging who supplies its most expensive input. Qualcomm will provide custom AI-inference silicon and optical connectivity for AWS, which widens Amazon's options beyond its own Trainium and Nvidia. The detail worth holding is the payment structure: Amazon receives milestone-linked Qualcomm warrants, so it collects equity in its supplier as purchases accumulate. That is a buyer using its scale to take up to 25M shares of the seller, and it beats a discount because it pays off if Qualcomm's AI business works for everyone else too.

SPCX: SpaceX Class A rose 2.2% to $151.21 and remains 6.1% below its Jun 12 first-day close and 28% under its $211 high. Starlink is compounding fast: subscriptions doubled to 12M at Jun 30 from 6M a year earlier. The supply overhang is an underwriting agreement releasing holdings in 7% tranches measured against holder-specific eligible balances.

AVGO: Broadcom added 1.1% to $361.99 and is still 7.9% below its Aug 14 close and 25% under its $482 high, which is a strange place for a company that just reported $16.7B of quarterly AI-semiconductor revenue, up 221%. Free cash flow hit $13.7B, 46% of total revenue, a conversion rate most hardware businesses never see. Management guided to $21.7B of AI-semiconductor revenue next quarter. All of that landed on Sep 2, before this week, so the 1.1% gain is a partial recovery against an outlook the market already holds. The question is not whether demand exists but whether shipments convert on schedule. Broadcom designs custom accelerators for a small number of very large customers, which makes its revenue lumpy and its order book concentrated. Hitting the $21.7B guide confirms the conversion; missing it while free cash flow stays near 46% of revenue would say the business is fine and the timing is not. At 25% off the high, the market is already pricing some of that slippage.

META: Meta rose 5.1% to $648.03, the best week among the mega caps, and it is 9.9% above its Aug 14 close. Muse, announced Sep 8, is the reason: task execution bolted onto the largest consumer distribution network in the world, free at entry with paid tiers for heavy use. The economics cut both ways and Meta has been unusually clear about it. Muse runs on a dedicated virtual machine, so every completed task carries a compute bill, and Meta states that Muse conversations and virtual-machine data do not reach its advertising systems. That closes off the obvious monetization shortcut. This launch has to earn its infrastructure through subscriptions, not better ad targeting, which is a harder and more honest business.

TSLA: Tesla rallied 3.2% to $365.44 and is still 10.1% below its Jun 12 close and 25% under its $490 high. The company sells two things to two different sets of believers. The vehicle and storage business runs on deliveries, prices and incentives, and its last reported quarter ended Jun 30, which makes the current tape an eleven-week-old fundamental with a story on top. The autonomy business runs on operating permissions, paid deployments and cost per mile, and none of those show up in a delivery number. With the quarter closing, a strong delivery print only helps if it arrives with evidence of the margin retained after incentives, because volume bought with discounts is the version that flatters the headline and misses the earnings. Deliveries up with stable margin confirms demand; deliveries up with margin compression means Tesla bought the quarter. Storage remains the quiet third exposure almost nobody prices.

ORCL: Oracle fell 5.4% to $150.28 and the week's most striking number is not the growth, it is the funding. Infrastructure revenue rose 121% to $7.4B for the quarter ended Aug 31, and capital spending still exceeded operating cash flow by $5.4B, even though $11.4B of customer prepayments was already sitting inside that operating line. Oracle covered the gap with a common-stock sale raising $19.9B net. So the fastest-growing cloud business is being financed by its customers and its shareholders at the same time, and the shareholders are paying in dilution. The stock is 54% below its $328 high, the deepest drawdown among the large caps here.

PLTR: Palantir fell 4.1% to $167.23, and its Nebius partnership is the kind of announcement that reads bigger than it is. The two will offer a sovereign-AI stack combining Palantir's controlled deployment environment with Nebius compute, aimed at customers whose data cannot leave a defined boundary. What the announcement does not contain is a revenue split, a minimum purchase, or a timetable, which means neither company has committed anything measurable. Palantir gets wider platform adoption if it works; Nebius gets paid for infrastructure.

HOOD: Robinhood sank 7.8% to $112.57 after gaining 17.1% the week before. Read it beside crypto, not beside the brokers: August crypto volume rose 61% month over month while options contracts fell 10% and event contracts fell 23%. Customers rotated toward crypto and Robinhood Chain, away from the products Robinhood monetizes most reliably. Net deposits slowed to $4B from $5.6B in July, which is the number that funds every other line. At 26% below its high and 73% above its low, this is the most cyclical equity on the page.

ADBE: Adobe fell 5.4% to $252.23 in the week it raised full-year guidance. Revenue grew 13% to $6.76B and AI-first annual recurring revenue grew more than 150%, a number Adobe declined to attach a dollar base to. The billion monthly active users has the same problem: it counts people, not payers. What the market can price is the fourth-quarter midpoint, which implies roughly 1% sequential growth, and $2.52B of operating cash flow that proves the legacy franchise still prints. So the tape read a 13% grower guiding to 1% sequential and marked it 31% below its $367 high.

LITE: Lumentum rose 5.2% to $927.03 and is flat against Aug 14, which undersells what has happened: the stock is up 520% from its 52-week low of $149.61 and sits 12% below its high. Optical components turned out to be a bottleneck in AI data centers, and the market repriced the whole category. The plan was public before this week: Aug 11 guidance put fiscal first-quarter revenue at a $1.25B midpoint, 24.2% above the prior quarter, with a 39.5% to 40.5% non-GAAP operating margin. That pairing is the interesting part, because a 24% sequential ramp usually costs margin and Lumentum says it will not. Supply is the constraint that decides it: the AXT agreement locks indium-phosphide wafers years out, and orders only become revenue when qualified components ship. Hitting the $1.25B midpoint with margin inside the guided band confirms the ramp is supply-secured; margin slipping below 39.5% means the company bought volume with cost. After a 520% run, execution is the entire thesis.

CRWV: CoreWeave slipped 0.4% to $88.99, and that calm weekly print hides a 15.5% fall since Aug 14 and a 38% drawdown from its $143 high. This is a levered balance sheet wearing a cloud-computing label. CoreWeave finances GPU capacity against customer contracts, so shareholders carry three risks at once: demand for computing, the cost of the debt, and the credit of the customers backing it. The May 18 disclosure makes the terms concrete: a $3.1B delayed-draw facility priced at SOFR plus 4.5 percentage points, serving two non-investment-grade customers. At current short rates that is expensive money lent against counterparties who cannot borrow cheaply themselves.

COIN: Coinbase fell 5.1% to $175.26 and is 18% above its Aug 14 close. The week delivered a real regulatory step: a Sep 8 SEC notice confirmed Coinbase Derivatives' Sep 1 filing to register as a national securities exchange for security futures under Section 6(g). Registration under 6(g) becomes effective on the notice filing itself, so this is not a discretionary approval sitting in a queue. What it does not do is bless any particular product: no single-stock perpetual has been approved or launched on the back of it. Coinbase now owns the venue rights and still has to build the market. The stock sits 55% below its $387 high, the deepest drawdown on this page.

KR: Kroger ended 0.2% lower at $58.49 after rising 2.7% on Friday. Identical sales excluding fuel rose just 0.2%, and inside that number is a 138-basis-point drag tied to the Inflation Reduction Act. Management cut the full-year identical-sales range to 0.2% to 0.8% and left EPS and free-cash-flow guidance untouched. Lowering the sales forecast while holding the profit forecast is a statement that the earnings no longer depend on the store. Retail media and e-commerce are doing that work: advertising carries margins a grocery aisle cannot approach. The stock is 5.3% above its 52-week low, so the market is not paying for the mix shift yet. EPS holding while identical sales land at the bottom of the range confirms the advertising offset; EPS following sales down means the grocery business still sets the earnings and the mix story is decoration.

CPRT: Copart fell 11.2% to $29.95 and changed strategy in the same week its profitability broke. Fourth-quarter revenue grew 2.4% while gross profit fell 5.5% and net income fell 17.4%: growth at the top, contraction everywhere that counts. Copart runs salvage-vehicle auctions, a business whose economics depend on volume per facility and take rate, both of which are now moving the wrong way. The answer is a $1.9B cash acquisition of dealer-auction marketplace ACV at $10.50 a share, and management expects EPS accretion only from fiscal 2028. So investors are asked to fund an expansion beyond the core franchise while the core franchise is shrinking, and to wait two years for the arithmetic. At 38% below its high, some of that is priced.

CASY: Casey's plunged 18.6% to $615.47 after reporting EPS of $7.37, up 27.7%, and the disconnect is the point. The earnings beat came from fuel margin expanding to 47.8 cents a gallon from 41.0, a 6.8-cent improvement, while same-store gallons fell 0.3%. Casey's makes money two ways: selling fuel at a 47.8-cent spread and selling food at a far wider one, and this quarter the spread did the work while volumes fell 0.3%. Inside-store traffic was positive, led by whole pizzas, which is the durable half of the business. Fuel margin is not: it widens when wholesale prices fall and compresses when they rise, and oil just gained 10.1% this week. Management held full-year guidance anyway. Down 27% from Aug 14, this is the worst four-week performance on this edition.

FICO: FICO rebounded 5.7% to $985.39 and recovered only a slice of the prior week's 19.2% collapse. The cause of both moves is the same Sep 9 Fannie Mae update: VantageScore 4.0 approved for all loan casefiles submitted or resubmitted under Desktop Underwriter 12.1. FICO licenses the score that has been the default in US mortgage underwriting for decades, and default status is the entire pricing power. Disclosed lender adoption of VantageScore at scale is the break; FICO holding price into next year's renewals with no measurable share loss confirms the moat survived the headline. At 48% below its $1,880 high and 6.8% above its low, the market has already decided which way this goes.

IREN: IREN eased 1.9% to $43.83 after a 26.0% gain the week before. The company reported $4B of contracted annual recurring revenue on Aug 27 against $70.5M of recognized AI revenue in the latest quarter. That is roughly 14 times the annualized run rate of what it currently books, and the contracted figure depends on delivery, commissioning, and specified prices that have not happened yet. Funding the gap costs real money: a $2.4B GPU financing package at a fixed 9% implies $216M of annual interest if fully drawn, before fees. So the company must commission capacity fast enough to outrun its own interest bill. Recognized AI revenue stepping up toward the contracted run rate over the next two quarters is the confirmation; the financing drawing down while recognized revenue stays near $70M is the break, and it is the same shape that has caught every debt-funded infrastructure build before this one.

LULU: Lululemon fell another 1.6% to $98.97 and is now 2.2% off its 52-week low and 54% below its high, the worst structural position on this report. The damage was done on Sep 3, when management guided to a 10% to 11% third-quarter revenue decline and 5% to 7% for the full year; the 16.7% drop came then, and this week was the aftershock. Second-quarter EPS of $2.92 is not the number to anchor on, because $0.86 of it came from tariff refunds and associated interest, so roughly 29% of reported earnings had nothing to do with selling clothes. The useful signal is inventory: units fell 7% while inventory dollars fell only 1%, meaning the average unit on hand got more expensive. That is either disciplined mix or unsold premium product, and sell-through decides which.

GME: GameStop jumped 10.4% to $21.15, and the operating story is finally outrunning the meme. Completed results on Sep 8 showed quarterly operating income more than doubling to $160.2M while sales fell 18.7% to $790.2M. Shrinking revenue with expanding profit means the mix changed, and it did: collectibles grew 57% and now make up 45.1% of sales. Trading cards carry margins a used-game shelf never did. The other half of this company is a portfolio, and it is large enough to matter more than the stores: $4.9B of eBay shares at quarter-end, with about $1.4B of convertible debt retired on Sep 3. Owning GME is owning a retailer and a concentrated equity position stapled together, and the eBay stake can swamp the operating result in either direction. The eBay stake sets the print. Until the stores outgrow it, GameStop trades as a holding company with a retailer attached.

CHWY: Chewy fell 13.6% to $20.44 in a week it raised full-year revenue and profitability guidance, which tells you the market was positioned for more. The quarter was solid: sales up 7.3% to $3.33B, or 5.7% excluding acquisitions; adjusted EBITDA margin improving to 6.8% from 5.9%. The number that defines this business is Autoship, which grew 9.3% and reached 84.6% of revenue. Nearly everything Chewy sells now arrives on a recurring schedule, which is why the margin keeps climbing: repeat delivery costs less to win than a first order. The unresolved question is where growth comes from. Customer count and spend per active customer separate expansion into new households from squeezing more out of existing ones, and only the first scales indefinitely. Active customers growing alongside spend confirms the former; spend rising while customer count stalls means Chewy is harvesting a fixed base, which caps the multiple. It sits 49% below its high.

TTAN: ServiceTitan collapsed 37.8% to $54.68 and closed the week exactly at its 52-week low, the single worst print in this edition. Revenue grew 21% to $292.8M with $50.5M of non-GAAP free cash flow, so the business is not broken. The guide is: $285M to $287M for next quarter, below the $292.8M just delivered. A software company guiding to sequential decline is telling the market its growth engine stalled, and that comparison is not seasonality-adjusted, which is the one caveat worth keeping. Management pushed the other story hard, saying it had exceeded its goal of doubling Max locations in Q2 and targeting over 700 enrolled locations by fiscal year-end.

AVAV: AeroVironment added 1.4% to $146.71 and remains 64% below its $410 high. The quarter was better than the chart: revenue up 6% to $480.5M, a 1.4 book-to-bill, and funded backlog up 37% to $1.5B. Book-to-bill above one means the company sold more than it delivered, which replenishes the work available for future quarters, and a 37% backlog build is a genuine demand signal in defense procurement. Gross margin improved to 26% from 21%, though the honest reading is that most of the gain came from lower intangible amortization and purchase-accounting costs rather than better manufacturing.

AXTI: AXT rose 5.1% to $64.77 and is still 20.7% below its Aug 14 close, which makes it the mirror image of Lumentum, its own customer, sitting flat over the same stretch. The link is contractual: a July agreement to supply indium-phosphide wafers to Lumentum through 2031, with an initial $43.5M deposit due within 30 business days and a second $43.5M payment whose timing and terms get set in 2028. The stock is up more than sixteen-fold from its 52-week low, so the market has already paid for the contract working.

BRZE: Braze slumped 25.1% to $23.94 after a quarter that looked good everywhere except the guide. Revenue grew 26.2% to $227.2M, non-GAAP operating income jumped to $22M from $6M, free cash flow reached $21.7M, and dollar-based net retention improved to 110% from 108%. That last number is the one that usually holds software multiples up: existing customers spent 10% more than a year ago, and the trend is improving. Then management guided next quarter to a $229.5M midpoint, about 1% above the quarter just reported. Growing 26% and guiding to 1% sequential is the same shape that broke ServiceTitan this week, and the market applied the same discount. Braze sells customer-engagement software, so its revenue tracks how much its clients are messaging their own users, which makes it a second derivative of consumer marketing budgets. Management's Sep 15 investor appearance is the dated opportunity to explain the gap

NBIS: Nebius fell 0.8% to $224.55. Its news is the other side of Palantir's: a sovereign-AI partnership letting eligible commercial customers reach Nebius compute and inference inside Palantir's controlled environment. For Nebius this is distribution, and distribution is the scarce input for a challenger cloud, because compute is a commodity and access to regulated customers is not. The agreement carries no minimum purchase and no revenue timetable. Nebius supplies infrastructure; Palantir supplies the environment and the customer relationship, which means Palantir controls the margin if this works. It trades 22% below its high and roughly three times its 52-week low; the market is still pricing the challenger case.

QCOM: Qualcomm rose 7.8% to $181.97, the best equity print of the week. The Amazon collaboration covering custom inference silicon and optical connectivity is a genuine entry into AI infrastructure for a company the market still values as a mobile-chip designer. Qualcomm disclosed warrants for up to 25M shares at an initial $161.26 exercise price, with Amazon initially holding vested rights over 3.75M. More rights vest as purchases hit milestones, governed by a $60B payment threshold that Qualcomm did not describe as firm backlog. So the customer is paid in Qualcomm equity for buying Qualcomm products, and the cheaper the shares are when exercised, the more of the company the customer owns. That is the cost of entry into this market. The stock is 28% below its high.

Commodities: The Price of Getting Oil Sooner
The commodity complex split between the one market that is short of supply and everything else. October WTI gained 10.1% to $100.62 while October Henry Hub fell 3.9% and December gold, silver and copper all declined. Grain moved on its own calendar, with USDA cutting its 2026/27 US corn production forecast by 213M bushels to 15.8B on a 178.5-bushel yield and trimming feed and residual use by 150M bushels, which limited the ending-stock reduction to 5.2%. Oil was not merely the best performer. It was the only market where the curve said scarcity.


Friday's preliminary CME settlements put October WTI at $100.05 and April 2027 at $80.11, a $19.94 premium for covering an immediate need over securing spring delivery. Buyers pay that gap in cash to get barrels now. Behind it sits a specific interruption: Saudi Arabia's precautionary shutdown of the East-West pipeline on Sep 11, with the IEA still counting more than 10M barrels a day of Gulf output shut in through August and global observed inventories down 95M barrels on the month. Seven OPEC+ producers left October quotas unchanged on Sep 6, which told the market nothing about barrels available next week.

US natural gas is the mirror image. EIA put Lower 48 working stocks at 3,254 Bcf on Sep 4, which is 148 Bcf above the five-year average, down from the 160 Bcf surplus reported in the previous edition and 79 Bcf below last year. The cushion narrowed 12 Bcf in a week, in the direction that matters heading into heating season, even as EIA's winter outlook still expects end-October stocks 5% above their five-year average on Permian and Haynesville production. Distillate tells the same story as oil rather than gas. Stocks rose 2.1M barrels with refineries running at 97.8% of capacity and still sit 13% below the five-year average, at a latest level of 106.3M barrels against 120.6M a year ago.

The metals all lost ground to the same competitor. December gold fell 2.0% to $4,391 as the ten-year real yield rose to 2.60%, and it fell despite $18B of August ETF inflows and central banks adding about 130 tonnes through July, down from 160 tonnes a year earlier. Silver fell harder at 2.9% with net longs already thinning to 26,049 contracts, while copper fell 1.8% even as net longs rose 11,607 to 92,476. Copper is the one metal still within 3.8% of its 52-week high, which makes it the cleanest read on industrial demand in a week when financial demand lost to real rates.




Gold's fund buyers arrived before the rate move that punished them. Global physically backed funds held 4,068 tonnes on Jul 31 and added 121 tonnes in August to reach a record 4,189 tonnes on Aug 31. Then December gold fell 2.0% this week as the ten-year real yield rose 17 basis points to 2.60%. Record holdings met a rising competing return, and the return won. That sequence is the thing to remember about gold: the fund flow tells you what allocators did last month, and the real yield tells you what the price will do next.

XAU: December gold fell 2.0% to $4,391 in the same month funds put $18B into it. That gap is the whole read. Gold pays nothing, so its price is mostly a statement about the return on the thing that does, and the ten-year real yield rose 17 basis points to 2.60% this week. Real yields are the competing asset, and they won. Underneath, the official bid is cooling: central banks bought about 130 tonnes through July against 160 tonnes a year earlier, a 19% slowdown in the one buyer that never sells on price. Gold sits 17% below its $5,318 high and 0.7% above its Aug 14 close, so this is a stall, not a break. Real yields turning back below 2.60% hands the trade straight back to gold; another quarter of official buying at 130 tonnes says the floor everyone assumes is under this market is softer than advertised. Own gold here and you are short real rates, whatever the central-bank says.

XAG: Silver lost 2.9% to $64.91, deeper than gold's 2.0%, and it is the most damaged chart in the complex at 43% below its $115.08 high. Silver trades like levered gold with an industrial job attached, which is why it falls harder in the same environment. Positioning had already thinned before the drop: CFTC put noncommercial net longs at 26,049 contracts on Sep 8, down 690 on the week. The pair trade is the position: silver gives you gold's rate exposure with a factory cycle bolted on.

WTI: Oil is the one market paying up for time. October WTI gained 10.1% to $100.62, while April 2027 settled at $80.11. That $19.94 gap is the whole story: crude at Cushing is not scarce, prompt barrels are. Saudi Arabia confirmed a precautionary shutdown of the East-West pipeline on Sep 11, the same artery Aramco's Aug 4 results credited with keeping exports moving through the Hormuz disruption. The IEA still has more than 10M barrels a day of Gulf output shut in and global inventories down 95M barrels on the month. Seven OPEC+ producers left October quotas unchanged on Sep 6, which says nothing about barrels available next week, and the market priced that correctly. Refiners and fuel buyers pay this premium in cash. A confirmed East-West restart with restored throughput collapses the spread and the trade with it; the shutdown statement carried no restart date, which is why the premium exists. If the gap is still above $15 at October expiry, this is a physical shortage rather than a precaution. WTI sits 21% above its Aug 14 close and 11% under the $112.95 high, so the squeeze has room left in it.

XCU: Copper fell 1.8% to $6.55 and is still the strongest chart in the commodity complex, 3.8% under its $6.80 high while silver sits 43% below its own and gas 62% below. Traders leaned into that: CFTC noncommercial net longs rose 11,607 contracts to 92,476 on Sep 8, the mirror image of silver's cut in the same report. So positioning grew 14% while the price fell 1.8%, which is the setup that resolves violently in one direction or the other. Copper is the industrial read the rest of the complex is missing, and the contract quoted here is December COMEX, US delivery, not LME cash. Price reclaiming $6.68 validates the build; longs liquidating from 92,476 without a demand catalyst turns a crowded position into the supply. Copper is the most direct expression of the physical-economy trade.

NG: US natural gas fell 3.9% to $2.826, and its cushion is quietly thinning. EIA put Sep 4 working stocks 148 Bcf above the five-year average, down from the 160 Bcf surplus reported in the previous edition and 79 Bcf below last year. That is a 12 Bcf narrowing in a single week, in the direction that matters before heating season. Gas is the anti-oil trade this week: oil is paying a $19.94 premium for prompt barrels while gas buyers sit on a seasonal buffer that Permian and Haynesville production keeps refilling. Another two weekly prints shrinking the surplus toward zero before end-October turns a comfortable winter into a tight one; EIA's forecast holding stocks 5% above average at end-October keeps the buffer intact and the price capped. At 62% below its $7.46 high, gas is priced for the comfortable version.

URA: URA fell 5.5% to $43.53 and the interesting part is what it did not do. Miner-focused URNM dropped 8.2% in the same week. The difference is structural: Global X holds 22.4% in Cameco and 6.5% in the Sprott physical trust, so a slice of the fund is stored uranium rather than equity, and stored fuel does not have a cost curve to miss. That blend cushioned the fall, and it is the reason to own this wrapper instead of the pure miners. The demand side remains unconverted. Cameco said at the symposium it had sent term sheets to prospective new buyers and that none had bought uranium yet. A signed contract off those term sheets is the catalyst the whole sector is waiting on; term sheets aging into the fourth quarter without a signature says the utility bid is talk. Both against SPY's 0.8% loss, which is the benchmark this fund has to beat to justify the volatility.

FCX: Freeport fell 2.3% to $71.07 against December copper's 1.8% decline, which means the equity gave you almost nothing the metal did not. That is the problem with owning the miner for copper exposure right now. FCX is copper with an operating lever attached, and this week the lever did not move.

Conclusion
Two prices did the work this week and neither came from a central bank. A pipeline shutdown in Saudi Arabia put a $19.94 premium on prompt crude, and real yields rose 17 basis points to 2.60% before the inflation data arrived. Everything else on this page was sorted by exposure to those two costs: the names that sell something scarce or collect a growing fee held up, and the names financing a build, renting distant cash flows, or burning fuel did not.
Three things resolve in the next ten days. The Fed decides on Sep 16, with Sep 16 retail sales landing the same day and covering August spending that predates the confidence drop households reported this month. The East-West pipeline either restarts with confirmed throughput, which collapses the prompt premium, or it does not, which turns a precaution into a physical shortage. And Micron reports Sep 30, the first clean read on whether the AI budget is still growing or merely changing hands.
Watch the spreads rather than the levels. RSP catching up to SPX would say the weakness was rotation. ETH funds gathering for a second week would make this a genuine allocation shift rather than a holiday-week air pocket. Oil's October-to-April gap holding above $15 into expiry would say the supply story has outlived the headline that started it.


