Agents, memory and the 6 trillion mark: the full chip week in five videos

Agents are rewriting the data center

With a classic chatbot the job was simple: ask a question, the graphics processor generates the reply, and the central processor keeps the workload balanced. Agents change the picture because an agent receives a goal rather than a single task; it splits the goal into steps, decides at each step what to research, which tool to call, which code to write and which model to consult, then checks its work and loops until the goal is met. Reasoning and planning run on the graphics chip, while tool calls, code execution and database lookups largely run on the CPU. That shift is dragging the data-center GPU-to-CPU ratio from 8-to-1 toward 4, then 2, and in some setups toward more CPUs than GPUs. The ratio shift brought scarcity with it. AMD chief Lisa Su said the server CPU supply chain has stayed tight since early in the year, with much of the demand never forecasted. Intel management likewise admitted it can serve only about half of its customers. A week after that admission, AMD shares jumped nearly 10 percent in a day and crossed the 1 trillion dollar market value for the first time. The agent wave, in other words, is not only a software trick; it is a physical reshuffling of rack contents that rewrites the revenue math of chipmakers.

The memory crunch overshadows the CPU crunch

The second bottleneck gets at least as much airtime, and the picture is harsher on the memory side. Micron expects tight memory conditions to extend beyond 2027, while the message from Samsung points to shortages lasting through 2028 with supply trailing demand. The company says it can cover only half to two-thirds of demand from major customers. Prices reacted fast: DRAM rose more than 60 percent in the spring, with talk of five to sevenfold jumps in some memory lines. PCGamer's industry roundup independently confirms the squeeze described in the video, reporting that Samsung fills only about 70 percent of orders and that cautious spending discipline at the two giants sets the stage for the crunch to persist into 2028.

The presenter's preference for Micron over AMD rests on a universality argument. AMD reaching 46 percent of the x86 server market by revenue is a major feat, yet the company fights Intel and also Arm-based designs plus Nvidia's custom Vera silicon. Each extra point of share is harder than the last because it means selling into customers embedded in rival ecosystems. Memory works like a neutral levy: whoever wins the GPU or CPU contest, DRAM sits next to every processor. The memory maker never has to bet on the architecture war's winner; the war itself keeps filling its coffers.

The clearest arithmetic behind that argument sits inside Nvidia's Vera Rubin racks. Each rack holds 36 superchips, pairing one Vera CPU with two Rubin GPUs. Every GPU carries 288 GB of high-bandwidth memory and every CPU up to 1.5 TB of low-power DRAM. That adds up to almost 21 TB on the graphics side and as much as 54 TB on the CPU side, so as agents multiply, most of the memory entering the rack lands beside the CPU, not the GPU. CoreWeave's production measurements back the architecture: on Vera Rubin NVL72, tokens per megawatt run up to 10 times the GB200 figure, and a 20.7 TB unified HBM4 domain turns a single rack into a giant inference engine.

A second mechanism feeding memory appetite is the model's working record, the KV cache . As a model advances through a task it keeps a running log of everything so far, precisely so it need not reprocess the past for each new token. Longer sessions inflate that log; once the GPU's high-bandwidth memory fills, the overflow spills first into plain DRAM and then into solid-state drives. Because agent sessions run far longer than classic chats, the spillover is turning from exception into routine. The result is an extra layer of memory demand beyond the rising CPU count.

The scale and multiple gap

The scale figures are striking: Micron's data-center business now brings in nearly three times the combined data-center revenue of AMD and Intel, according to the presenter, with Samsung's memory sales higher still. Both companies sit near the trillion-dollar band while Wall Street expects each to roughly double earnings next year. The invoice differs: investors pay over 45 times expected earnings for AMD and under seven times for Micron, making AMD about seven times pricier per dollar of expected profit. Zacks data relayed via theglobeandmail.com confirms the gap, with Micron at 5.81 and AMD at 39.27 times forward earnings.

AMD's latest quarter explains why the multiple is lofty. Revenue hit a record 11.5 billion dollars, up 50 percent, with the data-center business more than doubling on Instinct GPUs and Epyc server CPU sales up over 70 percent. Guidance near 13 billion dollars for next quarter comes with server revenue seen up more than 80 percent in the year's second half and over 70 percent on top of that in 2027. Two giant deals carry fine print worth reading: OpenAI and Meta hold warrants to buy 160 million shares each for a penny apiece, potentially owning up to 16 percent of the company after dilution as more GPUs ship. The final price milestone sits at 600 dollars a share, already exceeded.

Micron's report a few days earlier tells a different story of scale and margin. As presented, the company beat its own forecast by about 8 percent with roughly five times year-ago revenue, producing four times AMD's sales at far fatter margins: 56 percent gross margin for AMD against 87 percent for Micron. Next-quarter guidance of 61.5 billion dollars implies a similar 13 percent climb from a base more than four times larger. Contracts anchor visibility: 26 long-term deals covering over a third of revenue through 2030 carry 32 billion dollars in customer commitments, and most high-bandwidth memory is sold out through 2027 at prices above this year's.

The risk ledger is as full as the glossy side. Memory prices keep climbing but the pace is cooling: after the spring surge above 60 percent, summer gains stayed below 20 percent, with management guiding slower increases ahead. Since much of the growth came from pricing, the company's own tempo is fading too, from 74 percent sequential growth two quarters ago to 31 percent last quarter and 13 percent expected next. The second risk is capital spending: 27 billion dollars went to fabs and equipment last fiscal year, with about 25 billion more planned in the next six months alone. Slowing prices plus swelling budgets point to margins drifting down over time, which explains half of Micron's discount; the other half sits in history.

Memory's brutal cycle is familiar to longtime holders. In spring 2022 the company posted 8.6 billion dollars in revenue, then slid to 3.7 billion nine months later as gross margins turned negative and nearly 2 billion dollars of chips were written down below production cost. Wafer output was cut around 20 percent, headcount 10 percent, and the stock halved. The valuation trap is classic: memory names look cheapest near the peak because earnings swell. At 4.5 times forward earnings in January 2018 the stock fell 27 percent within a year; at about 6.5 times in June 2022 it dropped another 35 percent by September. Under seven times again today, the shares tend to top before earnings do: in 2022 the stock peaked in January while revenue kept setting records into June.

What separates this cycle is high-bandwidth memory structurally devouring supply. HBM consumes roughly three times the silicon of plain DRAM, a ratio that climbs with every generation, so each new HBM wave drains supply that could have served ordinary memory. Price floors in Micron's long-term contracts cushion the ride, and the company expects margins to stay meaningfully above prior-cycle peaks even at those floors. Yet contracts cover only a third of revenue and the rest faces the market. A supply wave is also approaching: new fabs in Idaho and Japan start output between mid-2027 and late 2028, while Samsung and SK Hynix ramp too. Competition is heating: Samsung's HBM share leapt from 21 to 33 percent in one quarter while Micron's slipped from 21 to 18.

The verdict comes from asking what sits inside the price. Nvidia and Broadcom trade near 17 to 18 times forward earnings; at that multiple AMD would change hands under 250 dollars. The reverse math bites harder: to earn its quote at that multiple, AMD would need about 35 dollars a share next year against Wall Street's 15.60, before counting OpenAI and Meta dilution. Years of rapid growth are pre-loaded into the quote, and the company must keep gaining CPU and GPU share to justify it. Micron is the mirror: at 17 or 18 times, the shares would top 2700 dollars, so today's price only works if long-run earnings fall about 60 percent from next year's expectation and stay there. Analysts instead see earnings rising into 2028, the firm generated 33 billion dollars of free cash in a single quarter, holds 68 billion more cash than debt, and plans bigger buybacks. The plan is plain: keep buying Micron until DRAM contract prices flatten while fab spending continues, and on AMD wait for the dilution discount or an earnings blowout.

Peak, target and buyback triangle at Nvidia

On the Nvidia side the week's headline was the record: an intraday 237.83 dollars on October 2, clearing the May peak. The trigger was Morgan Stanley analyst Joseph Moore restoring Nvidia as his top semiconductor pick after meeting Jensen Huang and the finance chief, keeping an overweight rating and a 300-dollar target. The thesis has three legs: an early position in a fresh product cycle, durable demand from AI labs, and agent enthusiasm speeding Vera adoption. Management's roughly 70 percent revenue growth guide for fiscal 2028 reads conservative to the bank, which sees demand closer to 100 percent while Wall Street expected 45 to 50. According to Gurufocus's account of the note, the shares change hands near 15 times fiscal 2028 earnings, able to work through rising estimates without any multiple expansion.

The price looks familiar but the business underneath is far larger. The quarter behind the prior record pointed to about 68.1 billion dollars of revenue; fiscal Q2 2027 printed 96.2 billion, meaning quarterly sales grew 41 percent while the shares merely revisited the old band. The waypoint was an 81.6 billion first quarter, with yearly growth running 73, 85 and an extraordinary 106 percent. The engine is the data center: 62.3 billion to 75.2 billion to 89 billion, now over 90 percent of quarterly revenue. Nvidia Newsroom's official results confirm the picture, with data center up 117 percent on the year and 18 percent sequentially to 89 billion dollars, plus third-quarter guidance of 108 billion. Adjusted earnings per share rose about 40 percent from 1.59 to 2.22 dollars; sequential growth cooling from the 20s toward 12 percent remains remarkable at this scale.

That makes the valuation story cut the other way. Near 51 times earnings at the May peak, the shares now trade around 34 times at similar prices because profits outgrew the quote. The trailing four-quarter figure moved from about 4.64 to 6.95 dollars a share, roughly 50 percent higher. Two footnotes matter: recent GAAP results benefited from gains on stakes in other companies, which can flatter the classic multiple and reverse. The Street's average target sits in the 270s, with Morgan Stanley's 300 above it. The real trap is the bar itself: Nvidia can fall without missing, simply by delivering very good numbers that fall short of extraordinary. Four markers decide what follows: November's third quarter and its guidance, the Rubin and Vera timetable, hyperscaler capital spending, and the execution pace of the 235 billion buyback.

The buyback was the week's second Nvidia headline, and its size made history. On September 28 the board added 150 billion dollars to the program, lifting remaining authorization to 235 billion, billed as the largest authorization increase ever, for execution through fiscal 2028. The rationale, in Huang's words, is a once-in-a-generation platform shift plus strong cash generation. For scale, 235 billion is about 4 percent of a 5.5 trillion market value. Yahoo Finance calculates that spending it all requires roughly 44 billion dollars of repurchases a quarter through fiscal 2028, against 19.7 billion actually spent in Q2. The headline is generous; the cadence needs watching. Nvidia Newsroom's official release confirms the framework; what matters is how many shares leave the float at which prices and what that does to earnings per share.

Safety pivot: testing as the third system

The week's surprise headline came from safety rather than sales. The claim is blunt: the smartest people agree that self-willed, unbounded intelligence ends humanity, and politics will grasp it within 18 to 24 months and bar frontier training. Progress would not stop under a ban, only bend: instead of raising core intelligence, labs would field vast agent swarms that look like superintelligence to most people while remaining combined high-intelligence agents. Everyone from Amodei to Altman to Musk is said to voice the same fear in different tones. The video's spine is that Huang knows the train cannot be stopped, so he is laying track for it instead of standing in front.

That track is the Open Agent Safety Platform: the open-source OpenShell runtime boundary plus the Sentry reference design. OpenShell wraps the agent in a governed layer for files, tools and outside services, guarding credentials and logging permission calls, and it works with existing agents such as Claude Code and OpenAI Codex without rewrites. Sentry is the independent watcher running on BlueField-4 data processing units, quarantining an escaped agent within milliseconds. More than 100 organizations are on board, named as Microsoft, CrowdStrike, Cisco, Palo Alto Networks, Palantir, ServiceNow and Anthropic. Anthropic is adding protections to managed Claude agents while xAI uses the platform for coding agents and Grok models. CrowdStrike's technical note confirms the shape: OpenShell governs at the software layer while Sentry watches out-of-band through DOCA hardware, with the Falcon platform adding business context.

A television interview line from Huang reveals the commercial side: beside training and inference systems comes a third estate, environmental testing labs that prove products functional, capable and safe. Those testing estates will be enormous, since containing an intelligence of a given size demands a multiple of that capacity; the video cites a tenfold rule. Safety turns from luxury into requirement here: any company shipping agents without demonstrating best efforts stays fully exposed, a de-facto legal mandate even if no regulator writes one. Delayed lab listings are read as liability fear rather than profit trouble. With the shares climbing from 210 to 228 dollars and jumping after the September 28 reveal, the market may already be pricing that third cash stream. The TradingView news flow supports the read: more than 100 organizations and a 300-dollar target discussed in the same week.

What the charts and the options pit say

The chart desk speaks more cautiously. Nvidia closed up half a percent at 228 dollars, but a quarterly rebalance dumped the candle from the 230s in the final minutes. Three bearish candles now stand in a row, shooting star, bearish engulfing, shooting star, still needing confirmation under 227. That level is both yesterday's low and the fast-approaching 9-day average; holding it with closes in the 230s flips the picture bullish, and above 235 opens the path to 245, while losing it exposes the 225 gap, then 220 and the 217-218 band, with RSI in the mid-60s nearing overbought. The supporting cast looks stronger: AMD bounced off the 9-day line in hammer fashion, eyeing a trendline break toward retesting highs in the upper 600s, with closes under 600 and the high 590s instead inviting 570s-to-540s pressure. Broadcom plays the patience game under bearish daily averages, where 362 opens the low-to-mid 370s; TSMC is pristine on the weekly average though daily overbought conditions argue for breathing room toward 450 and the high 440s. The options pit joined the celebration: over 4 million contracts, more than a billion dollars in premium, four times more call money than put money, with market makers pricing near fifty-fifty odds of the 6 trillion line by month-end and over two-thirds by year-end. Quartz's market roundup (qz.com) completes the frame: up 27 percent in 2026 with a 5.7 trillion value under 300 billion short of the mark. Two sample trades package the view: a two-week 245 call calendar near 2.30 dollars profitable roughly between 237 and 254, and a 235-225-215 put butterfly risking 200 dollars for up to 800 if it pins 225.