xAI just undercut GPT-5.6 by half. Can it last?

TECH BRIEF

Hello and welcome to Thursday's brief, where the AI arms race showed up in five very different financial statements: SpaceX's xAI unit released Grok 4.6, a frontier model that matches OpenAI's GPT-5.6 Sol on the leading benchmark while undercutting its price by more than half, Google launched the Pixel 11 lineup with prices up as much as $100 and RAM trimmed in the base Pro models amid an industry-wide memory shortage, Cisco booked $4 billion in AI infrastructure orders from hyperscalers in a single quarter and raised its outlook, Bank of America committed $250 billion in financing to US data centers and energy infrastructure through mid-2027, and Tencent's AI spending surged 176%, squeezing profit even as revenue beat estimates.

• SpaceX's xAI releases Grok 4.6, matching GPT-5.6 Sol on the leading AI benchmark at less than half the price

• Google unveils the Pixel 11 lineup with prices up to $100 higher and less RAM in the base Pro models

• Cisco books $4 billion in AI infrastructure orders in a single quarter and raises its fiscal 2027 outlook

• Bank of America commits $250 billion to finance US data centers and energy infrastructure through July 2027

• Tencent's AI capital spending surges 176%, squeezing profit even as revenue beats estimates

Let's get into it.

SpaceX's xAI Ships Grok 4.6, Matching OpenAI's Best Model at Less Than Half the Price

• SpaceXAI, the AI unit under Elon Musk's SpaceX umbrella, released Grok 4.6 on Wednesday, scoring 61 on the Artificial Analysis Intelligence Index, tying OpenAI's GPT-5.6 Sol and landing one point behind Anthropic's Claude Fable 5

• Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens, unchanged from its predecessor Grok 4.5 and more than 60% cheaper than GPT-5.6 Sol's $5/$30 and Claude Opus 5's $5/$25

• The model topped several individual benchmarks, including GDPVal-AA v2, AA-Briefcase, and Harvey LAB, with xAI emphasizing long-running coding agents and more ambitious interactive and visual work

• Grok 4.6 is available through Cursor, Grok Build, a direct API, and infrastructure partners including OpenRouter, Vercel, and Cloudflare, with Cursor and Grok Build offering double usage limits for the first week

• Why it matters: xAI just matched the two best-funded labs in the industry on raw capability while undercutting them on price, intensifying a margin fight in frontier AI that neither OpenAI nor Anthropic can easily walk away from

SpaceXAI released Grok 4.6 on Wednesday, its first major model update since Grok 4.5. The model scored 61 on the Artificial Analysis Intelligence Index, a composite benchmark averaging results across nine tests, tying OpenAI's GPT-5.6 Sol for second place and finishing one point behind Anthropic's Claude Fable 5. Artificial Analysis found Grok 4.6 led outright on GDPVal-AA v2, AA-Briefcase, and Harvey LAB, benchmarks that lean toward professional and legal reasoning tasks. (VentureBeat, Artificial Analysis)

The pricing is the sharper part of the story. Grok 4.6 costs $2 per million input tokens and $6 per million output tokens, unchanged from Grok 4.5 and less than half of what GPT-5.6 Sol ($5/$30) or Claude Opus 5 ($5/$25) charge for comparable capability. xAI is emphasizing long-running coding agents and more ambitious interactive and visual work as the model's focus area, and it is shipping through Cursor, its own Grok Build product, a direct API, and infrastructure partners including OpenRouter, Vercel, and Cloudflare. (The Decoder)

Why it matters: Grok 4.6 lands at a moment when frontier labs are converging on similar capability levels, which pushes the competitive battle toward price and distribution rather than raw intelligence gains. xAI benefits from SpaceX's balance sheet and does not need frontier AI to be profitable on its own the way OpenAI and Anthropic do, giving it more room to undercut on price without matching margin pressure. For investors in the AI infrastructure trade, a credible entrant pricing aggressively below the two leaders is a reminder that inference margins across the sector remain far from settled, even as capital spending on the underlying compute keeps climbing.

Google Launches Pixel 11 Lineup With Higher Prices and Less RAM, Citing the Memory Shortage

• Google unveiled the Pixel 11, Pixel 11 Pro, Pixel 11 Pro XL, and Pixel 11 Pro Fold at a "Made by Google" event in New York on Wednesday, hosted by Trevor Noah, alongside the Pixel Watch 5 and a new $29 Pixel Tag tracker

• Prices rose across the line: the base Pixel 11 starts at $899, the Pro at $1,099, the Pro XL at $1,299, and the Pro Fold at $1,899, each roughly $100 above the equivalent 2025 model, while the Pixel Watch 5 rose $50 to $399

• The base Pixel 11 Pro and Pro XL ship with 12GB of RAM, down from 16GB in the prior generation, with a 16GB option available at extra cost, a change Google has tied to the industry-wide memory chip shortage

• All four phones use the new Tensor G6 chip, built on TSMC's 2nm process, which Google says is roughly 20% faster and 20% more power-efficient than the Pixel 10's Tensor G5, with a TPU 50% faster at running Gemini features

• Why it matters: the RAM cuts show a smartphone maker with Google's scale passing a global memory shortage directly to consumers rather than absorbing it, a preview of margin pressure likely to show up across the rest of the device industry this cycle

Google's Pixel 11 event unveiled a four-phone lineup alongside the Pixel Watch 5, updated Pixel Buds Pro 2, and a new $29 Pixel Tag tracker aimed at Apple's AirTag, along with a Quick Share update letting Pixel devices exchange files and contact info when tapped together. The Pixel 11 starts at $899 for 256GB, the Pixel 11 Pro at $1,099, the Pro XL at $1,299, and the Pro Fold at $1,899, each roughly $100 above last year's equivalent model, while the Pixel Watch 5 rose $50 to $399. All four phones run on the new Tensor G6 chip, Google's first Pixel processor built on TSMC's 2nm process, which the company says delivers 15% to 20% better efficiency than the Pixel 10's Tensor G5 along with a TPU that is 50% faster at processing Gemini commands. (Engadget)

The base Pixel 11 Pro and Pro XL ship with 12GB of RAM, down from 16GB in the equivalent Pixel 9 and Pixel 10 models, with a more expensive 16GB configuration available for buyers who want it. Google has attributed the cut to a severe, supplier-driven memory chip shortage that is also squeezing other device makers this cycle. (Engadget, PhoneArena)

Why it matters: Google is passing the cost of the ongoing memory shortage straight to buyers, through both higher sticker prices and a quiet spec downgrade on its higher-end phones, rather than absorbing the hit. That is a useful read on how much pricing power memory suppliers like Samsung, SK Hynix, and Micron currently hold, and it previews margin pressure that is likely to spread across PC and smartphone makers more broadly through the rest of this product cycle, adding a new variable to hardware names that have otherwise ridden the AI upgrade cycle higher.

Cisco Books $4 Billion in AI Orders in a Single Quarter, Raises Its Fiscal 2027 Outlook

• Cisco reported fiscal fourth-quarter revenue of $17.3 billion, with GAAP net income of $3.9 billion ($0.97 per share) and non-GAAP net income of $4.9 billion ($1.22 per share), capping a fiscal year in which full-year revenue rose 12% to $63.3 billion

• AI infrastructure orders from hyperscale customers reached $4 billion in the quarter alone, pushing full fiscal year 2026 AI orders to $9.3 billion, roughly 4.5 times what Cisco booked the prior year

• Total product orders grew 35% year over year in the quarter, or 25% excluding hyperscalers, with networking product orders up 40%, marking the eighth straight quarter of double-digit networking order growth

• Cisco guided fiscal 2027 revenue to roughly $72.8 billion, about 15% above fiscal 2026, and said it expects AI infrastructure revenue to climb 88% to $7.5 billion in fiscal 2027

• Why it matters: Cisco's order book is one of the cleanest read-throughs available on how much hyperscalers are actually spending on networking gear for AI data centers, and a 4.5-fold jump in a single year says that spending is still accelerating, not plateauing

Cisco's fourth-quarter results, reported Wednesday, closed out a fiscal year in which the company more than kept pace with the broader AI infrastructure buildout. Revenue for the quarter reached $17.3 billion, with GAAP net income of $3.9 billion and non-GAAP net income of $4.9 billion, while full fiscal 2026 revenue rose 12% to $63.3 billion with non-GAAP earnings of $4.33 per share for the year. (Cisco Newsroom)

The more telling number sits in the order book. AI infrastructure orders from hyperscale customers totaled $4 billion in the fourth quarter alone, bringing the full fiscal year total to $9.3 billion, about 4.5 times what Cisco booked in fiscal 2025. Total product orders climbed 35% year over year in the quarter, or 25% excluding hyperscalers, with networking product orders up 40% for an eighth consecutive quarter of double-digit growth. Cisco guided fiscal 2027 revenue to approximately $72.8 billion and said it expects AI infrastructure revenue specifically to grow 88% to $7.5 billion next fiscal year. (Unite.AI, Yahoo Finance)

Why it matters: unlike hyperscaler capex guidance, which is a forecast, Cisco's order book is money hyperscalers have already committed to spend on networking gear, making it one of the more concrete signals available on the pace of AI data center buildout. A 4.5-fold increase in AI orders over a single fiscal year, alongside guidance calling for continued acceleration into fiscal 2027, argues against the idea that hyperscaler AI infrastructure spending is nearing a plateau, at least on the networking side of the buildout that Cisco sees most directly.

Bank of America Commits $250 Billion to Finance the US Data Center and Energy Buildout

• Bank of America announced a "Critical Infrastructure Finance Initiative" on Wednesday, targeting $250 billion in lending, investment, and capital markets support for US infrastructure projects between January 2026 and July 2027

• The financing spans three areas: digital infrastructure including data centers and computing, energy and power infrastructure including renewable generation and storage, and core infrastructure such as transportation and natural gas

• The initiative combines primary market lending, direct investments, capital markets services, and banking and advisory work, positioning Bank of America alongside peers competing for financing mandates tied to the AI buildout

• The $250 billion target is measured over an 18-month window running from January 1, 2026 to July 4, 2027, one of the largest such commitments announced by a major US bank to date

• Why it matters: a quarter-trillion-dollar commitment from one of the largest US banks shows traditional lenders, not just private credit funds and sovereign wealth vehicles, are now underwriting the AI infrastructure buildout at scale

Bank of America unveiled its Critical Infrastructure Finance Initiative on Wednesday, pledging $250 billion in financing for US digital, energy, and core infrastructure projects over an 18-month window running from January 1, 2026 through July 4, 2027. The program will provide primary market lending, direct investments, capital markets services, and banking and advisory offerings, and the bank has framed it explicitly around rising demand tied to AI data centers, critical minerals, and energy infrastructure upgrades. (Yahoo Finance, Banking Dive)

The financing splits across three buckets: digital infrastructure such as data centers and computing capacity, energy and power infrastructure including renewable generation and storage, and core infrastructure like transportation and natural gas. The announcement puts Bank of America squarely alongside other large US banks that have been expanding their exposure to AI-linked infrastructure lending, a category that has grown rapidly as hyperscalers and their financing partners look beyond their own balance sheets to fund the buildout. (Fox Business)

Why it matters: most of the AI infrastructure financing structures reported so far, including the Google-Anthropic financing network covered in this column earlier this month, have run through private credit funds, special purpose vehicles, and equity warrants rather than traditional bank balance sheets. A commitment of this size from Bank of America signals that mainstream lenders now see the buildout as a durable, investable category rather than a niche risk, which widens the pool of capital available to fund it but also extends the exposure of the regulated banking system to a still-unproven AI capex cycle, a dynamic worth watching as more banks likely follow with similar programs.

Tencent's AI Spending Surges 176%, Squeezing Profit Even as Revenue Beats

• Tencent reported second-quarter revenue of RMB 204.8 billion (roughly $28.6 billion), up 11% year over year and ahead of the roughly RMB 202.2 billion analysts had expected

• Marketing services revenue rose 22% to RMB 43.6 billion as AI-powered tools let WeChat Mini-Shop advertisers automatically select products, generate creatives, and run smart bidding

• Capital expenditure surged 176% year over year to roughly RMB 52.8 billion (about $7.4 billion), pushing quarterly free cash flow negative by RMB 13.8 billion

• IFRS net profit grew just 0.7% year over year to RMB 56.0 billion, well below the roughly RMB 61.8 billion analysts expected, though core profit excluding new AI ventures rose 19% to RMB 86.1 billion

• Why it matters: Tencent is the clearest evidence yet that AI capex is now large enough to bend the profit line of even the most profitable Chinese internet companies, a pattern that mirrors what US hyperscalers have already shown investors this earnings season

Tencent reported second-quarter revenue of RMB 204.8 billion, up 11% year over year and ahead of the roughly RMB 202.2 billion consensus estimate, driven in large part by advertising. Marketing services revenue climbed 22% to RMB 43.6 billion as the company's AI Marketing Plus suite let WeChat Mini-Shop advertisers automatically select products, generate creatives, and run smart bidding, while gaming revenue also held up better than expected. (CNBC)

The spending side of the ledger is where the quarter turns. Capital expenditure surged 176% year over year to roughly RMB 52.8 billion, concentrated on AI compute capacity and data center buildout, which pushed quarterly free cash flow into negative territory at RMB 13.8 billion. Chairman and CEO Pony Ma Huateng said the spending reflects increased procurement of compute that the company expects to "convert usage of our applications and models into revenue going forward." IFRS net profit grew just 0.7% to RMB 56.0 billion, well short of the roughly RMB 61.8 billion analysts had modeled, though core profit excluding new AI ventures still rose 19% to RMB 86.1 billion. (Bloomberg, Cryptopolitan)

Why it matters: Tencent's negative quarterly free cash flow echoes what Alphabet reported earlier this month, when its own AI-related commitments pushed it to its first quarterly cash burn since 2004, suggesting AI capex is now large enough to bend the profit and cash flow lines of even the most dominant, most profitable internet platforms, not just capital-constrained challengers. For investors, the question raised by both the Tencent and Alphabet prints is the same: how long shareholders will tolerate profit growth trailing revenue growth before demanding evidence that the AI spending is converting into returns, and whether Chinese regulators or Tencent's own board pushes back on the pace before US peers do.